The field of computer graphics was revolutionized by models such as NeRF and 3D Gaussian Splatting, displacing triangles as the dominant representation for photogrammetry. In this paper, we argue for a triangle comeback. We develop a differentiable renderer that directly optimizes triangles via end-to-end gradients. We achieve this by rendering each triangle as differentiable splats, combining the efficiency of triangles with the adaptive density of representations based on independent primitives. Compared to popular 2D and 3D Gaussian Splatting methods, our approach achieves competitive rendering and convergence speed, and demonstrates high visual quality. On the Mip-NeRF360 dataset, our method outperforms concurrent nonvolumetric primitives in visual fidelity and achieves higher perceptual quality than the state-of-the-art Zip-NeRF on indoor scenes. Triangles are simple, compatible with standard graphics stacks and GPU hardware, and highly efficient. Our results highlight the efficiency and effectiveness of triangle-based representations for high-quality novel view synthesis. Triangles bring us closer to mesh-based optimization by combining classical computer graphics with modern differentiable rendering frameworks. The project page is https://trianglesplatting.github.io/

Image-based 3D object detection is widely employed in applications such as autonomous vehicles and robotics, yet current systems struggle with generalisation due to complex problem setup and limited training data. We introduce a novel pipeline that decouples 3D detection from 2D detection and depth prediction, using a diffusion-based approach to improve accuracy and support category-agnostic detection. Additionally, we introduce the Normalised Hungarian Distance (NHD) metric for an accurate evaluation of 3D detection results, addressing the limitations of traditional IoU and GIoU metrics. Experimental results demonstrate that our method achieves state-of-the-art accuracy and strong generalisation across various object categories and datasets.

Contemporary 3D research, particularly in reconstruction and generation, heavily relies on 2D images for inputs or supervision. However, current designs for these 2D-3D mapping are memory-intensive, posing a significant bottleneck for existing methods and hindering new applications. In response, we propose a pair of highly scalable components for 3D neural fields: Lightplane Renderer and Splatter, which significantly reduce memory usage in 2D-3D mapping (over$\mathbf{1 0 0 0} \times)$. These innovations enable the processing of vastly more and higher resolution images with significantly small memory and computational costs. We demonstrate their utility across various applications, from optimizing with image-level losses to enabling a versatile pipeline for scaling 3D reconstruction and generation.11work done during internship

3D scene generation has quickly become a challenging new research direction, fueled by consistent improvements of 2D generative diffusion models. Current methods generate scenes by iteratively stitching newly generated images with existing geometry, using pre-trained monocular depth estimators to lift the generated images to 3D. The predicted depth is fused with the existing scene representation through various alignment operations. In this work, we make two fundamental contributions to the field of 3D scene generation. First, we note that lifting images to 3D with a monocular depth estimation model is suboptimal as it ignores the geometry of the existing scene, thus prompting the need for alignment. We introduce a depth completion model to directly learn the 3D fusion process, resulting in improved geometric coherence of generated scenes. Second, we introduce a new benchmark to evaluate the geometric accuracy of scene generation methods. We show that the commonly used CLIP score between scene prompts and images is unsuitable for measuring the geometric quality of a scene and introduce a depth-based metric. Our benchmark thus offers an additional dimension to gauge the quality of generated scenes.

We propose Flash3D, a method for scene reconstruction and novel view synthesis from a single image which is both very generalisable and efficient. For generalisability, we start from a 'foundation' model for monocular depth estimation and extend it to a full 3D shape and appearance reconstructor. For efficiency, we base this extension on feed-forward Gaussian Splatting. Specifically, we predict a first layer of 3D Gaussians at the predicted depth, and then add additional layers of Gaussians that are offset in space, allowing the model to complete the reconstruction behind occlusions and truncations. Flash3D is very efficient, trainable on a single GPU in a day, and thus accessible to most researchers. It achieves state-of-the-art results when trained and tested on RealEstate10k. When transferred to unseen datasets like NYU it outperforms competitors by a large margin. More impressively, when transferred to KITTI, Flash3D achieves better PSNR than methods trained specifically on that dataset. In some instances, it even outperforms recent methods that use multiple views as input. Code, models, demo, and more results are available at https://www.robots.ox.ac.uk/~vgg/research/flash3d/.

Text- or image-to-3D generators and 3D scanners can now produce 3D assets with high-quality shapes and textures, but as single, fused entities lacking meaningful structure. In contrast, most applications and creative workflows require 3D assets to be composed of distinct, meaningful parts that can be independently manipulated. To bridge this gap, we introduce PartGen, a novel approach for generating, from text, images, or unstructured 3D objects, 3D objects composed of meaningful parts. Our method leverages a multi-view diffusion model to extract plausible and view-consistent part segmentations from multiple views of a 3D object, dividing it into meaningful components. A second multi-view diffusion model then processes each part individually, filling in occlusions and generating completed views, which are subsequently passed to a 3D reconstruction network. The completion process ensures that the reconstructed parts integrate cohesively by considering the context of the entire object, compensating for missing information caused by occlusions and, in extreme cases, hallucinating entirely invisible parts based on contextual cues. We evaluate PartGen on both generated and real 3D assets, demonstrating significant improvements over segmentation and part completion baselines. We also showcase downstream applications such as text-guided 3D part editing.

Recent advances in radiance field reconstruction, such as 3D Gaussian Splatting (3DGS), have achieved high-quality novel view synthesis and fast rendering by representing scenes with compositions of Gaussian primitives. However, 3D Gaussians present several limitations for scene reconstruction. Accurately capturing hard edges is challenging without significantly increasing the number of Gaussians, creating a large memory footprint. Moreover, they struggle to represent flat surfaces, as they are diffused in space. Without hand-crafted regularizers, they tend to disperse irregularly around the actual surface. To circumvent these issues, we introduce a novel method, named 3D Convex Splatting (3DCS), which leverages 3D smooth convexes as primitives for modeling geometrically-meaningful radiance fields from multi-view images. Smooth convex shapes offer greater flexibility than Gaussians, allowing for a better representation of 3D scenes with hard edges and dense volumes using fewer primitives. Powered by our efficient CUDA-based rasterizer, 3DCS achieves superior performance over 3DGS on benchmarks such as MipNeRF360, Tanks and Temples, and Deep Blending. Specifically, our method attains an improvement of up to 0.81 in PSNR and 0.026 in LPIPS compared to 3DGS while maintaining high rendering speeds and reducing the number of required primitives. Our results highlight the potential of 3D Convex Splatting to become the new standard for high-quality scene reconstruction and novel view synthesis. The project page is https://convexsplatting.github.io

The choice of data representation is a key factor in the success of deep learning in geometric tasks. For instance, DUSt3R recently introduced the concept of viewpoint-invariant point maps, generalizing depth prediction and showing that all key problems in the 3D reconstruction of static scenes can be reduced to predicting such point maps. In this paper, we develop an analogous concept for a very different problem: the reconstruction of the 3D shape and pose of deformable objects. To this end, we introduce Dual Point Maps (DualPM), where a pair of point maps is extracted from the same image—one associating pixels to their 3D locations on the object and the other to a canonical version of the object in its rest pose. We also extend point maps to amodal reconstruction to recover the complete shape of the object, even through self-occlusions. We show that 3D reconstruction and 3D pose estimation can be reduced to the prediction of DualPMs. Empirically, we demonstrate that this representation is a suitable target for deep networks to predict. Specifically, we focus on modeling quadrupeds, showing that DualPMs can be trained purely on synthetic 3D data, consisting of one or two models per category, while generalizing effectively to real images. With this approach, we achieve significant improvements over previous methods for the 3D analysis and reconstruction of such objects.

Temporal consistency is critical in video prediction to ensure that outputs are coherent and free of artifacts. Traditional methods, such as temporal attention and 3D convolution, may struggle with significant object motion and may not capture long-range temporal dependencies in dynamic scenes. To address this gap, we propose the Tracktention Layer, a novel architectural component that explicitly integrates motion information using point tracks, i.e., sequences of corresponding points across frames. By incorporating these motion cues, the Tracktention Layer enhances temporal alignment and effectively handles complex object motions, maintaining consistent feature representations over time. Our approach is computationally efficient and can be seamlessly integrated into existing models, such as Vision Transformers, with minimal modification. It can be used to upgrade image-only models to state-of-the-art video ones, sometimes outperforming models natively designed for video prediction. We demonstrate this on video depth prediction and video colorization, where models augmented with the Tracktention Layer exhibit significantly improved temporal consistency compared to baselines. Project website: zlai0.github.io/TrackTention.

We introduce Uncommon Objects in 3D (uCO3D), a new object-centric dataset for 3D deep learning and 3D generative AI. uCO3D is the largest publicly-available collection of high-resolution videos of objects with 3D annotations that ensures full-360° coverage. uCO3D is significantly more diverse than MVImgNet and CO3Dv2, covering more than 1,000 object categories. It is also of higher quality, due to extensive quality checks of both the collected videos and the 3D annotations. Similar to analogous datasets, uCO3D contains annotations for 3D camera poses, depth maps and sparse point clouds. In addition, each object is equipped with a caption and a 3D Gaussian Splat reconstruction. We train several large 3D models on MVImgNet, CO3Dv2, and uCO3D and obtain superior results using the latter, showing that uCO3D is better for learning applications.

Computer vision is largely based on 2D techniques, with 3D vision still relegated to a relatively narrow subset of applications. However, by building on recent advances in 3D models such as neural radiance fields, some authors have shown that 3D techniques can at last improve outputs extracted from independent 2D views, by fusing them into 3D and denoising them. This is particularly helpful in egocentric videos, where the camera motion is significant, but only under the assumption that the scene itself is static. In fact, as shown in the recent analysis conducted by EPIC Fields, 3D techniques are ineffective when it comes to studying dynamic phenomena, and, in particular, when segmenting moving objects. In this paper, we look into this issue in more detail. First, we propose to improve dynamic segmentation in 3D by fusing motion segmentation predictions from a 2D-based model into layered radiance fields (Layered Motion Fusion). However, the high complexity of long, dynamic videos makes it challenging to capture the underlying geometric structure, and, as a result, hinders the fusion of motion cues into the (incomplete) scene geometry. We address this issue through test-time refinement, which helps the model to focus on specific frames, thereby reducing the data complexity. This results in a synergy between motion fusion and the refinement, and in turn leads to segmentation predictions of the 3D model that surpass the 2D baseline by a large margin. This demonstrates that 3D techniques can enhance 2D analysis even for dynamic phenomena in a challenging and realistic setting.

We present VGGT, a feed-forward neural network that directly infers all key 3D attributes of a scene, including camera parameters, point maps, depth maps, and 3D point tracks, from one, a few, or hundreds of its views. This approach is a step forward in 3D computer vision, where models have typically been constrained to and specialized for single tasks. It is also simple and efficient, reconstructing images in under one second, and still outperforming alternatives that require post-processing with visual geometry optimization techniques. The network achieves state-of-the-art results in multiple 3D tasks, including camera parameter estimation, multi-view depth estimation, dense point cloud reconstruction, and 3D point tracking. We also show that using pretrained VGGT as a feature backbone significantly enhances downstream tasks, such as non-rigid point tracking and feed-forward novel view synthesis. Code and models are publicly available at https://github.com/facebookresearch/vggt.

We present Twinner, the first large reconstruction model capable of recovering a scene’s illumination as well as an object’s geometry and material properties from only a few posed images. Twinner is based on the Large Reconstruction Model and innovates in three key ways: 1)We introduce a memory-efficient voxel-grid transformer whose memory scales only quadratically with the size of the voxel grid. 2) To address the scarcity of high-quality ground-truth PBR- shaded models, we introduce a large, fully synthetic dataset of procedurally generated PBR-textured objects lit with varied illumination. 3) To bridge the synthetic-to-real gap, we finetune the model on real-world datasets using a differentiable physically based shading model, eliminating the need for ground-truth illumination or material properties, which are challenging to obtain in real-world scenarios. We demonstrate the efficacy of our model on the real-world StanfordORB benchmark, where, given a few input views, we achieve reconstruction quality significantly superior to existing feed-forward reconstruction networks and comparable to slower per-scene optimization methods.

We address the challenge of generating 3D worlds from textual descriptions. We propose SynCity, a training- and optimization-free approach, which leverages the geometric precision of pre-trained 3D generative models and the artistic versatility of 2D image generators to create large, high-quality 3D spaces. While most 3D generative models are object-centric and cannot generate large-scale worlds, we show how 3D and 2D generators can be combined to generate ever-expanding scenes. Through a tile-based approach, we allow fine-grained control over the layout and the appearance of scenes. The world is generated tile-by-tile, and each new tile is generated within its world-context and then fused with the scene. SynCity generates compelling and immersive scenes that are rich in detail and diversity.

We introduce Puppet-Master, an interactive video generator that captures the internal, part-level motion of objects, serving as a proxy for modeling object dynamics universally. Given an image of an object and a set of "drags" specifying the trajectory of a few points on the object, the model synthesizes a video where the object's parts move accordingly. To build Puppet-Master, we extend a pre-trained image-to-video generator to encode the input drags. We also propose all-to-first attention, an alternative to conventional spatial attention that mitigates artifacts caused by fine-tuning a video generator on out-of-domain data. The model is fine-tuned on Objaverse-Animation-HQ, a new dataset of curated part-level motion clips obtained by rendering synthetic 3D animations. Unlike real videos, these synthetic clips avoid confounding part-level motion with overall object and camera motion. We extensively filter sub-optimal animations and augment the synthetic renderings with meaningful drags that emphasize the internal dynamics of objects. We demonstrate that Puppet-Master learns to generate part-level motions, unlike other motion-conditioned video generators that primarily move the object as a whole. Moreover, Puppet-Master generalizes well to out-of-domain real images, outperforming existing methods on real-world benchmarks in a zero-shot manner.

DUSt3R has recently shown that one can reduce many tasks in multi-view geometry, including estimating camera intrinsics and extrinsics, reconstructing the scene in 3D, and establishing image correspondences, to the prediction of a pair of viewpoint-invariant point maps, i.e., pixel-aligned point clouds defined in a common reference frame. This formulation is elegant and powerful, but unable to tackle dynamic scenes. To address this challenge, we introduce the concept of Dynamic Point Maps (DPM), extending standard point maps to support 4D tasks such as motion segmentation, scene flow estimation, 3D object tracking, and 2D correspondence. Our key intuition is that, when time is introduced, there are several possible spatial and time references that can be used to define the point maps. We identify a minimal subset of such combinations that can be regressed by a network to solve the sub tasks mentioned above. We train a DPM predictor on a mixture of synthetic and real data and evaluate it across diverse benchmarks for video depth prediction, dynamic point cloud reconstruction, 3D scene flow and object pose tracking, achieving state-of-the-art performance. Code, models and additional results are available at https://www.robots.ox.ac.uk/~vgg/research/dynamic-point-maps/.

Generating high-quality 3D content from text, single images, or sparse view images remains a challenging task with broad applications. Existing methods typically employ multi-view diffusion models to synthesize multi-view images, followed by a feed-forward process for 3D reconstruction. However, these approaches are often constrained by a small and fixed number of input views, limiting their ability to capture diverse viewpoints and, even worse, leading to suboptimal generation results if the synthesized views are of poor quality. To address these limitations, we propose Flex3D, a novel two-stage framework capable of leveraging an arbitrary number of high-quality input views. The first stage consists of a candidate view generation and curation pipeline. In the second stage, the curated views are fed into a Flexible Reconstruction Model (FlexRM), built upon a transformer architecture that can effectively process an arbitrary number of inputs. Through extensive exploration of design and training strategies, we optimize FlexRM to achieve superior performance in both reconstruction and generation tasks. Our results demonstrate that Flex3D achieves state-of-the-art performance, with a user study winning rate of over 92% in 3D generation tasks when compared to several of the latest feed-forward 3D generative models.

We introduce AutoPartGen, a model that generates objects composed of 3D parts in an autoregressive manner. This model can take as input an image of an object, 2D masks of the object's parts, or an existing 3D object, and generate a corresponding compositional 3D reconstruction. Our approach builds upon 3DShape2VecSet, a recent latent 3D representation with powerful geometric expressiveness. We observe that this latent space exhibits strong compositional properties, making it particularly well-suited for part-based generation tasks. Specifically, AutoPartGen generates object parts autoregressively, predicting one part at a time while conditioning on previously generated parts and additional inputs, such as 2D images, masks, or 3D objects. This process continues until the model decides that all parts have been generated, thus determining automatically the type and number of parts. The resulting parts can be seamlessly assembled into coherent objects or scenes without requiring additional optimization. We evaluate both the overall 3D generation capabilities and the part-level generation quality of AutoPartGen, demonstrating that it achieves state-of-the-art performance in 3D part generation.

We present Farm3D, a method for learning category-specific 3D reconstructors for articulated objects, relying solely on “free” virtual supervision from a pre-trained 2D diffusion-based image generator. Recent approaches can learn a monocular network that predicts the 3D shape, albedo, illumination, and viewpoint of any object occurrence, given a collection of single-view images of an object category. However, these approaches heavily rely on manually curated clean training data, which are expensive to obtain. We propose a framework that uses an image generator, such as Stable Diffusion, to generate synthetic training data that are sufficiently clean and do not require further manual curation, enabling the learning of such a reconstruction network from scratch. Additionally, we incorporate the diffusion model as a score to enhance the learning process. The idea involves randomizing certain aspects of the reconstruction, such as viewpoint and illumination, generating virtual views of the reconstructed 3D object, and allowing the 2D network to assess the quality of the resulting image, thus providing feedback to the reconstructor. Unlike work based on distillation, which produces a single 3D asset for each textual prompt, our approach yields a monocular reconstruction network capable of outputting a controllable 3D asset from any given image, whether real or generated, in a single forward pass in a matter of seconds. Our network can be used for analysis, including monocular reconstruction, or for synthesis, generating articulated assets for real-time applications such as video games. The code can be found on the project page at https://farm3d.github.io/.

We propose a novel feed-forward 3D editing framework called Shap-editor. Prior research on editing 3D objects primarily concentrated on editing individual objects by leveraging off-the-shelf 2D image editing networks, utilizing a process called 3D distillation, which transfers knowl-edge from the 2D network to the 3D asset. Distillation ne-cessitates at least tens of minutes per asset to attain sat-isfactory editing results, thus it is not very practical. In contrast, we ask whether 3D editing can be carried out di-rectly by a feed-forward network, eschewing test-time op-timization. In particular, we hypothesise that this process can be greatly simplified by first encoding 3D objects into a suitable latent space. We validate this hypothesis by building upon the latent space of Shap-E. We demonstrate that direct 3D editing in this space is possible and efficient by learning a feed-forward editor network that only requires approximately one second per edit. Our experiments show that Shap-Editor generalises well to both in-distribution and out-of-distribution 3D assets with different prompts and achieves superior performance compared to methods that carry out test-time optimisation for each edited instance.

Advancements in 3D Gaussian Splatting have significantly accelerated 3D reconstruction and generation. However, it may require a large number of Gaussians, which creates a substantial memory footprint. This paper introduces GES (Generalized Exponential Splatting), a novel representation that employs Generalized Exponential Function (GEF) to model 3D scenes, requiring far fewer parti-cles to represent a scene and thus significantly outperforming Gaussian Splatting methods in efficiency with a plug-and-play replacement ability for Gaussian-based utilities. GES is validated theoretically and empirically in both principled 1D setup and realistic 3D scenes. It is shown to represent signals with sharp edges more accurately, which are typically challenging for Gaussians due to their inherent low-pass characteristics. Our empirical analysis demonstrates that GEF outperforms Gaussians in fitting natural-occurring signals (e.g. squares, triangles, parabolic signals), thereby reducing the need for extensive splitting operations that increase the memory footprint of Gaussian Splatting. With the aid of a frequency-modulated loss, GES achieves competitive performance in novel-view syn-thesis benchmarks while requiring less than half the memory storage of Gaussian Splatting and increasing the rendering speed by up to 39%. The code is available on the project website https://abdullahamdi.com/ges.

Learning 3D models of all animals in nature requires massively scaling up existing solutions. With this ultimate goal in mind, we develop 3D-Fauna, an approach that learns a pan-category deformable 3D animal model for more than 100 animal species jointly. One crucial bottle-neck of modeling animals is the limited availability of training data, which we overcome by learning our model from 2D Internet images. We show that prior approaches, which are category-specific, fail to generalize to rare species with limited training images. We address this challenge by introducing the Semantic Bank of Skinned Models (SBSM), which automatically discovers a small set of base animal shapes by combining geometric inductive priors with semantic knowledge implicitly captured by an off-the-shelf self-supervised feature extractor. To train such a model, we also contribute a new large-scale dataset of diverse animal species. At inference time, given a single image of any quadruped animal, our model reconstructs an articulated 3D mesh in a feed-forward manner in seconds.

We introduce the Splatter Image, an ultra-efficient approach for monocular 3D object reconstruction. Splatter Image is based on Gaussian Splatting, which allows fast and high-quality reconstruction of 3D scenes from multiple images. We apply Gaussian Splatting to monocular reconstruction by learning a neural network that, at test time, performs reconstruction in a feed-forward manner, at 38 FPS. Our main innovation is the surprisingly straightforward design of this network, which, using 2D operators, maps the input image to one 3D Gaussian per pixel. The resulting set of Gaussians thus has the form an image, the Splatter Image. We further extend the method take several images as input via cross-view attention. Owning to the speed of the renderer (588 FPS), we use a single GPU for training while generating entire images at each iteration to optimize perceptual metrics like LPIPS. On several synthetic, real, multi-category and large-scale benchmark datasets, we achieve better results in terms of PSNR, LPIPS, and other metries while training and evaluating much faster than prior works. Code, models and more results are available at https://szymanowiczs.github.io/splatter-image.

We introduce Free3D, a simple accurate method for monocular open-set novel view synthesis (NVS). Similar to Zero-1-to-S, we start from a pre-trained 2D image generator for generalization, and fine-tune it for NVS. Compared to other works that took a similar approach, we obtain significant improvements without resorting to an explicit 3D representation, which is slow and memory-consuming, and without training an additional network for 3D reconstruction. Our key contribution is to improve the way the target camera pose is encoded in the network, which we do by introducing a new ray conditioning normalization (RCN) layer. The latter injects pose information in the underlying 2D image generator by telling each pixel its viewing direction. We further improve multi-view consistency by using light-weight multi-view attention layers and by sharing generation noise between the different views. We train Free3D on the Objaverse dataset and demonstrate excellent generalization to new categories in new datasets, including OmniObject3D and GSO. The project page is available at https://chuanxiar.com/free3d/.

Most text-to-3D generators build upon off-the-shelf text-to-image models trained on billions of images. They use variants of Score Distillation Sampling (SDS), which is slow, somewhat unstable, and prone to artifacts. A mitigation is to fine-tune the 2D generator to be multi-view aware, which can help distillation or can be combined with reconstruction networks to output 3D objects directly. In this paper, we further explore the design space of text-to-3D models. We significantly improve multi-view generation by considering video instead of image generators. Combined with a 3D reconstruction algorithm which, by using Gaussian splatting, can optimize a robust image-based loss, we directly produce high-quality 3D outputs from the generated views. Our new method, IM-3D, reduces the number of evaluations of the 2D generator network 10-100$\times$, resulting in a much more efficient pipeline, better quality, fewer geometric inconsistencies, and higher yield of usable 3D assets.

We introduce MVSplat360, a feed-forward approach for 360° novel view synthesis (NVS) of diverse real-world scenes, using only sparse observations. This setting is inherently ill-posed due to minimal overlap among input views and insufficient visual information provided, making it challenging for conventional methods to achieve high-quality results. Our MVSplat360 addresses this by effectively combining geometry-aware 3D reconstruction with temporally consistent video generation. Specifically, it refactors a feed-forward 3D Gaussian Splatting (3DGS) model to render features directly into the latent space of a pre-trained Stable Video Diffusion (SVD) model, where these features then act as pose and visual cues to guide the denoising process and produce photorealistic 3D-consistent views. Our model is end-to-end trainable and supports rendering arbitrary views with as few as 5 sparse input views. To evaluate MVSplat360's performance, we introduce a new benchmark using the challenging DL3DV-10K dataset, where MVSplat360 achieves superior visual quality compared to state-of-the-art methods on wide-sweeping or even 360° NVS tasks. Experiments on the existing benchmark RealEstate10K also confirm the effectiveness of our model. Readers are highly recommended to view the video results at [donydchen.github.io/mvsplat360](https://donydchen.github.io/mvsplat360).

We consider the problem of segmenting objects in videos based on their motion and no other forms of supervision. Prior work has often approached this problem by using the principle of common fate, namely the fact that the motion of points that belong to the same object is strongly correlated. However, most authors have only considered instantaneous motion from optical flow. In this work, we present a way to train a segmentation network using long-term point trajectories as a supervisory signal to complement optical flow. The key difficulty is that long-term motion, unlike instantaneous motion, is difficult to model -- any parametric approximation is unlikely to capture complex motion patterns over long periods of time. We instead draw inspiration from subspace clustering approaches, proposing a loss function that seeks to group the trajectories into low-rank matrices where the motion of object points can be approximately explained as a linear combination of other point tracks. Our method outperforms the prior art on motion-based segmentation, which shows the utility of long-term motion and the effectiveness of our formulation.

We present Meta 3D AssetGen (AssetGen), a significant advancement in text-to-3D generation which produces faithful, high-quality meshes with texture and material control. Compared to works that bake shading in the 3D object’s appearance, AssetGen outputs physically-based rendering (PBR) materials, supporting realistic relighting. AssetGen generates first several views of the object with separate shaded and albedo appearance channels, and then reconstructs colours, metalness and roughness in 3D, using a deferred shading loss for efficient supervision. It also uses a sign-distance function to represent 3D shape more reliably and introduces a corresponding loss for direct shape supervision. This is implemented using fused kernels for high memory efficiency. After mesh extraction, a texture refinement transformer operating in UV space significantly improves sharpness and details. AssetGen achieves 17% improvement in Chamfer Distance and 40% in LPIPS over the best concurrent work for few-view reconstruction, and a human preference of 72% over the best industry competitors of comparable speed, including those that support PBR. Project page with generated assets: https://assetgen.github.io

Recent diffusion-based generators can produce high-quality images from textual prompts. However, they often disregard textual instructions that specify the spatial layout of the composition. We propose a simple approach that achieves robust layout control without the need for training or fine-tuning of the image generator. Our technique manipulates the cross-attention layers that the model uses to interface textual and visual information and steers the generation in the desired direction given, e.g., a user-specified layout. To determine how to best guide attention, we study the role of attention maps and explore two alternative strategies, forward and backward guidance. We thoroughly evaluate our approach on three benchmarks and provide several qualitative examples and a comparative analysis of the two strategies that demonstrate the superiority of backward guidance compared to forward guidance, as well as prior work. We further demonstrate the versatility of layout guidance by extending it to applications such as editing the layout and context of real images.

Most of us are not experts in specific fields, such as ornithology. Nonetheless, we do have general image and language understanding capabilities that we use to match what we see to expert resources. This allows us to expand our knowledge and perform novel tasks without ad-hoc external supervision. On the contrary, machines have a much harder time consulting expert-curated knowledge bases unless trained specifically with that knowledge in mind. Thus, in this paper we consider a new problem: fine-grained image recognition without expert annotations, which we address by leveraging the vast knowledge available in web encyclopedias. First, we learn a model to describe the visual appearance of objects using non-expert image descriptions. We then train a fine-grained textual similarity model that matches image descriptions with documents on a sentence-level basis. We evaluate the method on two datasets (CUB-200 and Oxford-102 Flowers) and compare with several strong baselines and the state of the art in cross-modal retrieval. Code is available at: https://github.com/subhc/clever.

Incremental object detection (IOD) aims to train an object detector in phases, each with annotations for new object categories. As other incremental settings, IOD is subject to catastrophic forgetting, which is often addressed by techniques such as knowledge distillation (KD) and exem-plar replay (ER). However, KD and ER do not work well if applied directly to state-of-the-art transformer-based object detectors such as Deformable DETR [59] and UP-DETR [9]. In this paper, we solve these issues by proposing a ContinuaL DEtection TRansformer (CL-DETR), a new method for transformer-based IOD which enables effective usage of KD and ER in this context. First, we introduce a Detector Knowledge Distillation (DKD) loss, focusing on the most informative and reliable predictions from old versions of the model, ignoring redundant background predictions, and ensuring compatibility with the available ground-truth labels. We also improve ER by proposing a calibration strategy to preserve the label distribution of the training set, therefore better matching training and testing statistics. We conduct extensive experiments on COCO 2017 and demonstrate that CL-DETR achieves state-of-the-art results in the IOD setting.11Code: https://1yy.mpi-inf.mpg.de/CL-DETR/

We introduce the novel-view acoustic synthesis (NVAS) task: given the sight and sound observed at a source view-point, can we synthesize the sound of that scene from an unseen target viewpoint? We propose a neural rendering approach: Visually-Guided Acoustic Synthesis (ViGAS) network that learns to synthesize the sound of an arbitrary point in space by analyzing the input audio-visual cues. To benchmark this task, we collect two first-of-their-kind large-scale multi-view audio-visual datasets, one synthetic and one real. We show that our model successfully reasons about the spatial cues and synthesizes faithful audio on both datasets. To our knowledge, this work represents the very first formulation, dataset, and approach to solve the novel-view acoustic synthesis task, which has exciting potential applications ranging from AR/VR to art and design. Unlocked by this work, we believe that the future of novel-view synthesis is in multi-modal learning from videos.

We consider the problem of reconstructing a dynamic scene observed from a stereo camera. Most existing meth-ods for depth from stereo treat different stereo frames in-dependently, leading to temporally inconsistent depth pre-dictions. Temporal consistency is especially important for immersive AR or VR scenarios, where flickering greatly di-minishes the user experience. We propose DynamicStereo, a novel transformer-based architecture to estimate dispar-ity for stereo videos. The network learns to pool information from neighboring frames to improve the temporal consistency of its predictions. Our architecture is designed to process stereo videos efficiently through divided attention layers. We also introduce Dynamic Replica, a new bench-mark dataset containing synthetic videos of people and ani-mals in scanned environments, which provides complemen-tary training and evaluation data for dynamic stereo closer to real applications than existing datasets. Training with this dataset further improves the quality of predictions of our proposed DynamicStereo as well as prior methods. Finally, it acts as a benchmark for consistent stereo methods. Project page: https://dynamic-stereo.github.io/

Diffusion models have emerged as the best approach for generative modeling of 2D images. Part of their success is due to the possibility of training them on millions if not billions of images with a stable learning objective. However, extending these models to 3D remains difficult for two reasons. First, finding a large quantity of 3D training data is much more complex than for 2D images. Second, while it is conceptually trivial to extend the models to operate on 3D rather than 2D grids, the associated cubic growth in memory and compute complexity makes this infeasible. We address the first challenge by introducing a new diffusion setup that can be trained, end-to-end, with only posed 2D images for supervision; and the second challenge by proposing an image formation model that decouples model memory from spatial memory. We evaluate our method on real-world data, using the CO3D dataset which has not been used to train 3D generative models before. We show that our diffusion models are scalable, train robustly, and are competitive in terms of sample quality and fidelity to existing approaches for 3D generative modeling.

Reconstructing the 3D shape of an object from a single RGB image is a long-standing problem in computer vision. In this paper, we propose a novel method for single-image 3D reconstruction which generates a sparse point cloud via a conditional denoising diffusion process. Our method takes as input a single RGB image along with its camera pose and gradually denoises a set of 3D points, whose positions are initially sampled randomly from a three-dimensional Gaussian distribution, into the shape of an object. The key to our method is a geometrically-consistent conditioning process which we call projection conditioning: at each step in the diffusion process, we project local image features onto the partially-denoised point cloud from the given camera pose. This projection conditioning process enables us to generate high-resolution sparse geometries that are well-aligned with the input image and can additionally be used to predict point colors after shape reconstruction. Moreover, due to the probabilistic nature of the diffusion process, our method is naturally capable of generating multiple different shapes consistent with a single input image. In contrast to prior work, our approach not only performs well on synthetic benchmarks but also gives large qualitative improvements on complex real-world data. Data and code are available at https://lukemelas.github.io/projectionconditioned-point-cloud-diffusion/.

We consider the problem of reconstructing a full 360° photographic model of an object from a single image of it. We do so by fitting a neural radiance field to the image, but find this problem to be severely ill-posed. We thus take an off-the-self conditional image generator based on diffusion and engineer a prompt that encourages it to “dream up” novel views of the object. Using the recent DreamFusion method, we fuse the given input view, the conditional prior, and other regularizers into a final, consistent reconstruction. We demonstrate state-of-the-art reconstruction results on benchmark images when compared to prior methods for monocular 3D reconstruction of objects. Qualitatively, our reconstructions provide a faithful match of the input view and a plausible extrapolation of its appearance and 3D shape, including to the side of the object not visible in the image.

Obtaining photorealistic reconstructions of objects from sparse views is inherently ambiguous and can only be achieved by learning suitable reconstruction priors. Earlier works on sparse rigid object reconstruction successfully learned such priors from large datasets such as CO3D. In this paper, we extend this approach to dynamic objects. We use cats and dogs as a representative example and introduce Common Pets in 3D (CoP3D), a collection of crowd-sourced videos of approximately 4,200 distinct pets. CoP3D is one of the first large-scale datasets for benchmarking non-rigid 3D reconstruction “in the wild”. We also propose Tracker-NeRF, a method for learning 4D reconstruction from our dataset. At test time, given a small number of video frames of an unseen object, Tracker-NeRF predicts the trajectories of its 3D points and generates new views, interpolating viewpoint and time. Results on CoP3D reveal significantly better non-rigid new-view synthesis performance than existing baselines. The data will be available on the project webpage: https://cop3d.github.io/.

We consider the problem of predicting the 3D shape, articulation, viewpoint, texture, and lighting of an articulated animal like a horse given a single test image as input. We present a new method, dubbed MagicPony, that learns this predictor purely from in-the-wild single-view images of the object category, with minimal assumptions about the topology of deformation. At its core is an implicit-explicit representation of articulated shape and appearance, combining the strengths of neural fields and meshes. In order to help the model understand an object's shape and pose, we distil the knowledge captured by an off-the-shelf self-supervised vision transformer and fuse it into the 3D model. To overcome local optima in viewpoint estimation, we further introduce a new viewpoint sampling scheme that comes at no additional training cost. MagicPony outperforms prior work on this challenging task and demonstrates excellent generalisation in reconstructing art, despite the fact that it is only trained on real images. The code can be found on the project page at https://3dmagicpony.github.io/

Diffusion-based image generators can now produce high-quality and diverse samples, but their success has yet to fully translate to 3D generation: existing diffusion methods can either generate low-resolution but 3D consistent outputs, or detailed 2D views of 3D objects but with potential structural defects and lacking view consistency or realism. We present HoloFusion, a method that combines the best of these approaches to produce high-fidelity, plausible, and diverse 3D samples while learning from a collection of multi-view 2D images only. The method first generates coarse 3D samples using a variant of the recently proposed HoloDiffusion generator. Then, it independently renders and upsamples a large number of views of the coarse 3D model, super-resolves them to add detail, and distills those into a single, high-fidelity implicit 3D representation, which also ensures view-consistency of the final renders. The super-resolution network is trained as an integral part of HoloFusion, end-to-end, and the final distillation uses a new sampling scheme to capture the space of super-resolved signals. We compare our method against existing baselines, including DreamFusion, Get3D, EG3D, and HoloDiffusion, and achieve, to the best of our knowledge, the most realistic results on the challenging CO3Dv2 dataset.

We introduce Replay, a collection of multi-view, multi-modal videos of humans interacting socially. Each scene is filmed in high production quality, from different view-points with several static cameras, as well as wearable action cameras, and recorded with a large array of microphones at different positions in the room. Overall, the dataset contains over 4000 minutes of footage and over 7 million timestamped high-resolution frames annotated with camera poses and partially with foreground masks. The Replay dataset has many potential applications, such as novel-view synthesis, 3D reconstruction, novel-view acoustic synthesis, human body and face analysis, and training generative models. We provide a benchmark for training and evaluating novel-view synthesis, with two scenarios of different difficulty. Finally, we evaluate several baseline state-of-the-art methods on the new benchmark.

Large-scale Vision-Language Models, such as CLIP, learn powerful image-text representations that have found numerous applications, from zero-shot classification to text-to-image generation. Despite that, their capabilities for solving novel discriminative tasks via prompting fall behind those of large language models, such as GPT-3. Here we explore the idea of visual prompt engineering for solving computer vision tasks beyond classification by editing in image space instead of text. In particular, we discover an emergent ability of CLIP, where, by simply drawing a red circle around an object, we can direct the model’s attention to that region, while also maintaining global information. We show the power of this simple approach by achieving state-of-the-art in zero-shot referring expressions comprehension and strong performance in keypoint localization tasks. Finally, we draw attention to some potential ethical concerns of large language-vision models.

We present Viewset Diffusion, a diffusion-based generator that outputs 3D objects while only using multi-view 2D data for supervision. We note that there exists a one-to-one mapping between viewsets, i.e., collections of several 2D views of an object, and 3D models. Hence, we train a diffusion model to generate viewsets, but design the neural network generator to reconstruct internally corresponding 3D models, thus generating those too. We fit a diffusion model to a large number of viewsets for a given category of objects. The resulting generator can be conditioned on zero, one or more input views. Conditioned on a single view, it performs 3D reconstruction accounting for the ambiguity of the task and allowing to sample multiple solutions compatible with the input. The model performs reconstruction efficiently, in a feed-forward manner, and is trained using only rendering losses using as few as three views per viewset. Project page: szymanowiczs.github.io/viewset-diffusion.

Vector Quantisation (VQ) is experiencing a comeback in machine learning, where it is increasingly used in representation learning. However, optimizing the codevectors in existing VQ-VAE is not entirely trivial. A problem is codebook collapse, where only a small subset of codevectors receive gradients useful for their optimisation, whereas a majority of them simply "dies off" and is never updated or used. This limits the effectiveness of VQ for learning larger codebooks in complex computer vision tasks that require high-capacity representations. In this paper, we present a simple alternative method for online codebook learning, Clustering VQ-VAE (CVQ-VAE). Our approach selects encoded features as anchors to update the "dead" codevectors, while optimising the codebooks which are alive via the original loss. This strategy brings unused codevectors closer in distribution to the encoded features, increasing the likelihood of being chosen and optimized. We extensively validate the generalization capability of our quantiser on various datasets, tasks (e.g. reconstruction and generation), and architectures (e.g. VQ-VAE, VQGAN, LDM). CVQ-VAE can be easily integrated into the existing models with just a few lines of code.

We present MAV3D (Make-A-Video3D), a method for generating three-dimensional dynamic scenes from text descriptions. Our approach uses a 4D dynamic Neural Radiance Field (NeRF), which is optimized for scene appearance, density, and motion consistency by querying a Text-to-Video (T2V) diffusion-based model. The dynamic video output generated from the provided text can be viewed from any camera location and angle, and can be composited into any 3D environment. MAV3D does not require any 3D or 4D data and the T2V model is trained only on Text-Image pairs and unlabeled videos. We demonstrate the effectiveness of our approach using comprehensive quantitative and qualitative experiments and show an improvement over previously established internal baselines. To the best of our knowledge, our method is the first to generate 3D dynamic scenes given a text description. Generated samples can be viewed at make-a-video3d.github.io

Instance segmentation in 3D is a challenging task due to the lack of large-scale annotated datasets. In this paper, we show that this task can be addressed effectively by leveraging instead 2D pre-trained models for instance segmentation. We propose a novel approach to lift 2D segments to 3D and fuse them by means of a neural field representation, which encourages multi-view consistency across frames. The core of our approach is a slow-fast clustering objective function, which is scalable and well-suited for scenes with a large number of objects. Unlike previous approaches, our method does not require an upper bound on the number of objects or object tracking across frames. To demonstrate the scalability of the slow-fast clustering, we create a new semi-realistic dataset called the Messy Rooms dataset, which features scenes with up to 500 objects per scene. Our approach outperforms the state-of-the-art on challenging scenes from the ScanNet, Hypersim, and Replica datasets, as well as on our newly created Messy Rooms dataset, demonstrating the effectiveness and scalability of our slow-fast clustering method.

Neural rendering is fuelling a unification of learning, 3D geometry and video understanding that has been waiting for more than two decades. Progress, however, is still hampered by a lack of suitable datasets and benchmarks. To address this gap, we introduce EPIC Fields, an augmentation of EPIC-KITCHENS with 3D camera information. Like other datasets for neural rendering, EPIC Fields removes the complex and expensive step of reconstructing cameras using photogrammetry, and allows researchers to focus on modelling problems. We illustrate the challenge of photogrammetry in egocentric videos of dynamic actions and propose innovations to address them. Compared to other neural rendering datasets, EPIC Fields is better tailored to video understanding because it is paired with labelled action segments and the recent VISOR segment annotations. To further motivate the community, we also evaluate two benchmark tasks in neural rendering and segmenting dynamic objects, with strong baselines that showcase what is not possible today. We also highlight the advantage of geometry in semi-supervised video object segmentations on the VISOR annotations. EPIC Fields reconstructs 96\% of videos in EPIC-KITCHENS, registering 19M frames in 99 hours recorded in 45 kitchens, and is available from: http://epic-kitchens.github.io/epic-fields

In this paper we tackle the problem of Generalized Category Discovery (GCD). Specifically, given a dataset with labelled and unlabelled images, the task is to cluster all images in the unlabelled subset, whether or not they belong to the labelled categories. Our first contribution is to recognise that most existing GCD benchmarks only contain labels for a single clustering of the data, making it difficult to ascertain whether models are leveraging the available labels to solve the GCD task, or simply solving an unsupervised clustering problem. As such, we present a synthetic dataset, named 'Clevr-4', for category discovery. Clevr-4 contains four equally valid partitions of the data, i.e based on object 'shape', 'texture' or 'color' or 'count'. To solve the task, models are required to extrapolate the taxonomy specified by labelled set, rather than simply latch onto a single natural grouping of the data. We use this dataset to demonstrate the limitations of unsupervised clustering in the GCD setting, showing that even very strong unsupervised models fail on Clevr-4. We further use Clevr-4 to examine the weaknesses of existing GCD algorithms, and propose a new method which addresses these shortcomings, leveraging consistent findings from the representation learning literature to do so. Our simple solution, which is based on `Mean Teachers' and termed $\mu$GCD, substantially outperforms implemented baselines on Clevr-4. Finally, when we transfer these findings to real data on the challenging Semantic Shift Benchmark suite, we find that $\mu$GCD outperforms all prior work, setting a new state-of-the-art.

Video provides us with the spatio-temporal consistency needed for visual learning. Recent approaches have utilized this signal to learn correspondence estimation from closeby frame pairs. However, by only relying on close-by frame pairs, those approaches miss out on the richer long-range consistency between distant overlapping frames. To address this, we propose a self-supervised approach for correspondence estimation that learns from multiview consistency in short RGB-D video sequences. Our approach combines pairwise correspondence estimation and registration with a novel SE(3) transformation synchronization algorithm. Our key insight is that self-supervised multiview registration allows us to obtain correspondences over longer time frames; increasing both the diversity and difficulty of sampled pairs. We evaluate our approach on indoor scenes for correspondence estimation and RGB-D pointcloud registration and find that we perform on-par with supervised approaches.

Abstract Learning deformable 3D objects from 2D images is often an ill-posed problem. Existing methods rely on explicit supervision to establish multi-view correspondences, such as template shape models and keypoint annotations, which restricts their applicability on objects “in the wild”. A more natural way of establishing correspondences is by watching videos of objects moving around. In this paper, we present DOVE, a method that learns textured 3D models of deformable object categories from monocular videos available online, without keypoint, viewpoint or template shape supervision. By resolving symmetry-induced pose ambiguities and leveraging temporal correspondences in videos, the model automatically learns to factor out 3D shape, articulated pose and texture from each individual RGB frame, and is ready for single-image inference at test time. In the experiments, we show that existing methods fail to learn sensible 3D shapes without additional keypoint or template supervision, whereas our method produces temporally consistent 3D models, which can be animated and rendered from arbitrary viewpoints. Project page: https://dove3d.github.io/ .

We propose a method to learn 3D deformable object categories from raw single-view images, without external supervision. The method is based on an autoencoder that factors each input image into depth, albedo, viewpoint and illumination. In order to disentangle these components without supervision, we use the fact that many object categories have, at least approximately, a symmetric structure. We show that reasoning about illumination allows us to exploit the underlying object symmetry even if the appearance is not symmetric due to shading. Furthermore, we model objects that are probably, but not certainly, symmetric by predicting a symmetry probability map, learned end-to-end with the other components of the model. Our experiments show that this method can recover very accurately the 3D shape of human faces, cat faces and cars from single-view images, without any supervision or a prior shape model. On benchmarks, we demonstrate superior accuracy compared to another method that uses supervision at the level of 2D image correspondences.

We present Neural Feature Fusion Fields (N3F), a method that improves dense 2D image feature extractors when the latter are applied to the analysis of multiple images reconstructible as a 3D scene. Given an image feature extractor, for example pre-trained using self-supervision, N3F uses it as a teacher to learn a student network defined in 3D space. The 3D student network is similar to a neural radiance field that distills said features and can be trained with the usual differentiable rendering machinery. As a consequence, N3F is readily applicable to most neural rendering formulations, including vanilla NeRF and its extensions to complex dynamic scenes. We show that our method not only enables semantic understanding in the context of scene-specific neural fields without the use of manual labels,but also consistently improves over the self-supervised 2D baselines. This is demonstrated by considering various tasks, such as 2D object retrieval, 3D segmentation, and scene editing, in diverse sequences, including long egocentric videos in the EPIC-KITCHENS benchmark. Project page: https://www.robots.ox.ac.uk/~vadim/n3f/.

We tackle the problem of learning object detectors without supervision. Differently from weakly-supervised object detection, we do not assume image-level class labels. Instead, we extract a supervisory signal from audio-visual data, using the audio component to “teach” the object detector. While this problem is related to sound source localisation, it is considerably harder because the detector must classify the objects by type, enumerate each instance of the object, and do so even when the object is silent. We tackle this problem by first designing a self-supervised framework with a contrastive objective that jointly learns to classify and localise objects. Then, without using any supervision, we simply use these self-supervised labels and boxes to train an image-based object detector. With this, we outperform previous unsupervised and weakly-supervised detectors for the task of object detection and sound source localization. We also show that we can align this detector to ground-truth classes with as little as one label per pseudo-class, and show how our method can learn to detect generic objects that go beyond instruments, such as airplanes and cats.

Unsupervised localization and segmentation are long-standing computer vision challenges that involve decom-posing an image into semantically meaningful segments without any labeled data. These tasks are particularly interesting in an unsupervised setting due to the difficulty and cost of obtaining dense image annotations, but existing un-supervised approaches struggle with complex scenes containing multiple objects. Differently from existing methods, which are purely based on deep learning, we take inspiration from traditional spectral segmentation methods by re-framing image decomposition as a graph partitioning problem. Specifically, we examine the eigenvectors of the Laplacian of a feature affinity matrix from self-supervised networks. We find that these eigenvectors already decompose an image into meaningful segments, and can be readily used to localize objects in a scene. Furthermore, by clustering the features associated with these segments across a dataset, we can obtain well-delineated, nameable regions, i.e. semantic segmentations. Experiments on complex datasets (PASCAL VOC, MS-COCO) demonstrate that our simple spectral method outperforms the state-of-the-art in unsupervised localization and segmentation by a significant margin. Furthermore, our method can be readily usedfor a variety of complex image editing tasks, such as background removal and compositing.11Project Page: https://lukemelas.github.io/deep-spectral-segmentation/

We consider the problem of reconstructing the depth of dynamic objects from videos. Recent progress in dynamic video depth prediction has focused on improving the output of monocular depth estimators by means of multi-view constraints while imposing little to no restrictions on the deformation of the dynamic parts of the scene. However, the theory of Non-Rigid Structure from Motion prescribes to constrain the deformations for 3D reconstruction. We thus propose a new model that departs significantly from this prior work. The idea is to fit a dynamic point cloud to the video data using Sinkhorn's algorithm to associate the 3D points to 2D pixels and use a differentiable point renderer to ensure the compatibility of the 3D deformations with the measured optical flow. In this manner, our algorithm, called Keypoint Transporter, models the overall deformation of the object within the entire video, so it can constrain the reconstruction correspondingly. Compared to weaker deformation models, this significantly reduces the reconstruction ambiguity and, for dynamic objects, allows Keypoint Transporter to obtain reconstructions of the quality superior or at least comparable to prior approaches while being much faster and reliant on a pre-trained monocular depth estimator network. To assess the method, we evaluate on new datasets of synthetic videos depicting dynamic humans and animals with ground-truth depth. We also show qualitative results on crowd-sourced real-world videos of pets.

In this paper, we consider a highly general image recognition setting wherein, given a labelled and unlabelled set of images, the task is to categorize all images in the unlabelled set. Here, the unlabelled images may come from labelled classes or from novel ones. Existing recognition methods are not able to deal with this setting, because they make several restrictive assumptions, such as the unlabelled instances only coming from known – or unknown – classes, and the number of unknown classes being known a-priori. We address the more unconstrained setting, naming it ‘Generalized Category Discovery’, and challenge all these assumptions. We first establish strong baselines by taking state-of-the-art algorithms from novel category discovery and adapting them for this task. Next, we propose the use of vision transformers with contrastive representation learning for this open-world setting. We then introduce a simple yet effective semi-supervised k-means method to cluster the unlabelled data into seen and unseen classes automatically, substantially outperforming the baselines. Finally, we also propose a new approach to estimate the number of classes in the unlabelled data. We thoroughly evaluate our approach on public datasets for generic object classification and on fine-grained datasets, leveraging the recent Semantic Shift Benchmark suite. Code: https://www.robots.ox.ac.uk/~vgg/research/gcd

Prior work for articulated 3D shape reconstruction often relies on specialized multi-view and depth sensors or pre-built deformable 3D models. Such methods do not scale to diverse sets of objects in the wild. We present a method that requires neither of them. It aims to create high-fidelity, articulated 3D models from many casual RGB videos in a differentiable rendering framework. Our key in-sight is to merge three schools of thought: (1) classic deformable shape models that make use of articulated bones and blend skinning, (2) canonical embeddings that establish correspondences between pixels and a canonical 3D model, and (3) volumetric neural radiance fields (NeRFs) that are amenable to gradient-based optimization. We introduce neural blend skinning models that allow for differentiable and invertible articulated deformations. When combined with canonical embeddings, such models allow us to establish dense correspondences across videos that can be self-supervised with cycle consistency. On real and synthetic datasets, our method shows higher-fidelity 3D reconstructions than prior works for humans and animals, with the ability to render realistic images from novel viewpoints. Project page: https://banmo-www.github.io/.

We present a system for automatic converting of 2D mask object predictions and raw LiDAR point clouds into full 3D bounding boxes of objects. Because the LiDAR point clouds are partial, directly fitting bounding boxes to the point clouds is meaningless. Instead, we suggest that obtaining good results requires sharing information between all objects in the dataset jointly, over multiple frames. We then make three improvements to the baseline. First, we address ambiguities in predicting the object rotations via direct optimization in this space while still backpropagating rotation prediction through the model. Second, we explicitly model outliers and task the network with learning their typical patterns, thus better discounting them. Third, we enforce temporal consistency when video data is available. With these contributions, our method significantly outperforms previous work despite the fact that those methods use significantly more complex pipelines, 3D models and additional human-annotated external sources of prior information.

We propose a new approach to learn to segment multiple image objects without manual supervision. The method can extract objects form still images, but uses videos for supervision. While prior works have considered motion for segmentation, a key insight is that, while motion can be used to identify objects, not all objects are necessarily in motion: the absence of motion does not imply the absence of objects. Hence, our model learns to predict image regions that are likely to contain motion patterns characteristic of objects moving rigidly. It does not predict specific motion, which cannot be done unambiguously from a still image, but a distribution of possible motions, which includes the possibility that an object does not move at all. We demonstrate the advantage of this approach over its deterministic counterpart and show state-of-the-art unsupervised object segmentation performance on simulated and real-world benchmarks, surpassing methods that use motion even at test time. As our approach is applicable to variety of network architectures that segment the scenes, we also apply it to existing image reconstruction-based models showing drastic improvement. Project page and code: https://www.robots.ox.ac.uk/~vgg/research/ppmp.

We tackle the problem of discovering novel classes in an image collection given labelled examples of other classes. We present a new approach called AutoNovel to address this problem by combining three ideas: (1) we suggest that the common approach of bootstrapping an image representation using the labelled data only introduces an unwanted bias, and that this can be avoided by using self-supervised learning to train the representation from scratch on the union of labelled and unlabelled data; (2) we use ranking statistics to transfer the model's knowledge of the labelled classes to the problem of clustering the unlabelled images; and, (3) we train the data representation by optimizing a joint objective function on the labelled and unlabelled subsets of the data, improving both the supervised classification of the labelled data, and the clustering of the unlabelled data. Moreover, we propose a method to estimate the number of classes for the case where the number of new categories is not known a priori. We evaluate AutoNovel on standard classification benchmarks and substantially outperform current methods for novel category discovery. In addition, we also show that AutoNovel can be used for fully unsupervised image clustering, achieving promising results.

Differently from 2D image datasets such as COCO, largescale human datasets with 3D ground-truth annotations are very difficult to obtain in the wild. In this paper, we address this problem by augmenting existing 2D datasets with high-quality 3D pose fits. Remarkably, the resulting annotations are sufficient to train from scratch 3D pose regressor networks that outperform the current state-of-the-art on in the-wild benchmarks such as 3DPW. Additionally, training on our augmented data is straightforward as it does not require to mix multiple and incompatible 2D and 3D datasets or to use complicated network architectures and training procedures. This simplified pipeline affords additional improvements, including injecting extreme crop augmentations to better reconstruct highly truncated people, and incorporating auxiliary inputs to improve 3D pose estimation accuracy. It also reduces the dependency on 3D datasets such as H36M that have restrictive licenses. We also use our method to introduce new benchmarks for the study of real-world challenges such as occlusions, truncations, and rare body poses. In order to obtain such high quality 3D pseudo-annotations, inspired by progress in internal learning, we introduce Exemplar Fine-Tuning (EFT). EFT combines the re-projection accuracy of fitting methods like SMPLify with a 3D pose prior implicitly captured by a pre-trained 3D pose regressor network. We show that EFT produces 3D annotations that result in better downstream performance and are qualitatively preferable in an extensive human-based assessment.

Given a raw video sequence taken from a freely-moving camera, we study the problem of decomposing the observed 3D scene into a static background and a dynamic foreground containing the objects that move within the scene. This task is reminiscent of the classic background subtraction problem, but is significantly harder because all parts of the scene, static and dynamic, generate a large apparent motion due to the camera large viewpoint change and parallax. In particular, we consider egocentric videos and further separate the dynamic component into objects and the actor that observes and moves them. We achieve this factorization by reconstructing the video via a triple-stream neural rendering network that explains the different motions based on corresponding inductive biases. We demonstrate that our method can successfully separate the different types of motion, outperforming recent neural rendering baselines at this task, and can accurately segment the moving objects. We do so by assessing the method empirically on challenging videos from the EPIC-KITCHENS dataset which we augment with appropriate annotations to create a new benchmark for the task of dynamic object segmentation on unconstrained video sequences, for complex 3D environments.

The objective of this work is to localize sound sources that are visible in a video without using manual annotations. Our key technical contribution is to show that, by training the network to explicitly discriminate challenging image fragments, even for images that do contain the object emitting the sound, we can significantly boost the localization performance. We do so elegantly by introducing a mechanism to mine hard samples and add them to a contrastive learning formulation automatically. We show that our algorithm achieves state-of-the-art performance on the popular Flickr SoundNet dataset. Furthermore, we introduce the VGG-Sound Source (VGG-SS) benchmark, a new set of annotations for the recently-introduced VGG-Sound dataset, where the sound sources visible in each video clip are explicitly marked with bounding box annotations. This dataset is 20 times larger than analogous existing ones, contains 5K videos spanning over 200 categories, and, differently from Flickr SoundNet, is video-based. On VGG-SS, we also show that our algorithm achieves state-of-the-art performance against several baselines. Code and datasets can be found at http://www.robots.ox.ac.uk/˜vgg/research/lvs/.

We present NeuroMorph, a new neural network architecture that takes as input two 3D shapes and produces in one go, i.e. in a single feed forward pass, a smooth interpolation and point-to-point correspondences between them. The interpolation, expressed as a deformation field, changes the pose of the source shape to resemble the target, but leaves the object identity unchanged. NeuroMorph uses an elegant architecture combining graph convolutions with global feature pooling to extract local features. During training, the model is incentivized to create realistic deformations by approximating geodesics on the underlying shape space manifold. This strong geometric prior allows to train our model end-to-end and in a fully unsupervised manner without requiring any manual correspondence annotations. NeuroMorph works well for a large variety of input shapes, including non-isometric pairs from different object categories. It obtains state-of-the-art results for both shape correspondence and interpolation tasks, matching or surpassing the performance of recent unsupervised and supervised methods on multiple benchmarks.

Our goal is to learn a deep network that, given a small number of images of an object of a given category, reconstructs it in 3D. While several recent works have obtained analogous results using synthetic data or assuming the avail-ability of 2D primitives such as keypoints, we are interested in working with challenging real data and with no manual annotations. We thus focus on learning a model from multiple views of a large collection of object instances. We contribute with a new large dataset of object centric videos suitable for training and benchmarking this class of models. We show that existing techniques leveraging meshes, voxels, or implicit surfaces, which work well for reconstructing isolated objects, fail on this challenging data. Finally, we propose a new neural network design, called warp-conditioned ray embedding (WCR), which significantly improves reconstruction while obtaining a detailed implicit representation of the object surface and texture, also compensating for the noise in the initial SfM reconstruction that bootstrapped the learning process. Our evaluation demonstrates performance improvements over several deep monocular reconstruction baselines on existing benchmarks and on our novel dataset. For additional material please visit: https://henzler.github.io/publication/unsupervised_videos/.

Predicting future pedestrian trajectory is a crucial component of autonomous driving systems, as recognizing critical situations based only on current pedestrian position may come too late for any meaningful corrective action (e.g. breaking) to take place. In this paper, we propose a new method to predict future position of pedestrians, with respect to a predicted future position of the ego-vehicle, thus giving a assistive/autonomous driving system sufficient time to respond. The method explicitly disentangles actual movement of pedestrians in real world from the ego-motion of the vehicle, using a future pose prediction network trained in self-supervised fashion, which allows the method to observe and predict the intrinsic pedestrian motion in a normalised view, that captures the same real-world location across multiple frames.

We tackle the problem of learning the geometry of multiple categories of deformable objects jointly. Recent work has shown that it is possible to learn a unified dense pose predictor for several categories of related objects. However, training such models requires to initialize inter-category correspondences by hand. This is suboptimal and the resulting models fail to maintain correct correspondences as individual categories are learned. In this paper, we show that improved correspondences can be learned automatically as a natural byproduct of learning category-specific dense pose predictors. To do this, we express correspondences between different categories and between images and categories using a unified embedding. Then, we use the latter to enforce two constraints: symmetric inter-category cycle consistency and a new asymmetric image-to-category cycle consistency. Without any manual annotations for the intercategory correspondences, we obtain state-of-the-art alignment results, outperforming dedicated methods for matching 3D shapes. Moreover, the new model is also better at the task of dense pose prediction than prior work.

The quality of the image representations obtained from self-supervised learning depends strongly on the type of data augmentations used in the learning formulation. Recent papers have ported these methods from still images to videos and found that leveraging both audio and video signals yields strong gains; however, they did not find that spatial augmentations such as cropping, which are very important for still images, work as well for videos. In this paper, we improve these formulations in two ways unique to the spatio-temporal aspect of videos. First, for space, we show that spatial augmentations such as cropping do work well for videos too, but that previous implementations, due to the high processing and memory cost, could not do this at a scale sufficient for it to work well. To address this issue, we first introduce Feature Crop, a method to simulate such augmentations much more efficiently directly in feature space. Second, we show that as opposed to naïve average pooling, the use of transformer-based attention improves performance significantly, and is well suited for processing feature crops. Combining both of our discoveries into a new method, Space-Time Crop & Attend (STiCA) we achieve state-of-the-art performance across multiple video-representation learning benchmarks. In particular, we achieve new state-of-the-art accuracies of 67.0% on HMDB-51 and 93.1% on UCF-101 when pre-training on Kinetics-400. Code and pretrained models are available1.

In the image domain, excellent representations can be learned by inducing invariance to content-preserving transformations via noise contrastive learning. In this paper, we generalize contrastive learning to a wider set of transformations, and their compositions, for which either invariance or distinctiveness is sought. We show that it is not immediately obvious how existing methods such as SimCLR can be extended to do so. Instead, we introduce a number of formal requirements that all contrastive formulations must satisfy, and propose a practical construction which satisfies these requirements. In order to maximise the reach of this analysis, we express all components of noise contrastive formulations as the choice of certain generalized transformations of the data (GDTs), including data sampling. We then consider videos as an example of data in which a large variety of transformations are applicable, accounting for the extra modalities – for which we analyze audio and text – and the dimension of time. We find that being invariant to certain transformations and distinctive to others is critical to learning effective video representations, improving the state-of-the-art for multiple benchmarks by a large margin, and even surpassing supervised pretraining. Code and pretrained models are available1.

We tackle the problem of monocular 3D reconstruction of articulated objects like humans and animals. We contribute DensePose 3D, a method that can learn such reconstructions in a weakly supervised fashion from 2D image annotations only. This is in stark contrast with previous deformable reconstruction methods that use parametric models such as SMPL pre-trained on a large dataset of 3D object scans. Because it does not require 3D scans, DensePose 3D can be used for learning a wide range of articulated categories such as different animal species. The method learns, in an end-to-end fashion, a soft partition of a given category-specific 3D template mesh into rigid parts together with a monocular reconstruction network that predicts the part motions such that they reproject correctly onto 2D DensePose-like surface annotations of the object. The decomposition of the object into parts is regularized by expressing part assignments as a combination of the smooth eigenfunctions of the Laplace-Beltrami operator. We show significant improvements compared to state-of-the-art nonrigid structure-from-motion baselines on both synthetic and real data on categories of humans and animals.

Visuomotor control (VMC) is an effective means of achieving basic manipulation tasks such as pushing or pick- and-place from raw images. Conditioning VMC on desired goal states is a promising way of achieving versatile skill primitives. However, common conditioning schemes either rely on task-specific fine tuning - e.g. using one-shot imitation learning (IL) - or on sampling approaches using a forward model of scene dynamics i.e. model-predictive control (MPC), leaving deployability and planning horizon severely limited. In this paper we propose a conditioning scheme which avoids these pitfalls by learning the controller and its conditioning in an end-to-end manner. Our model predicts complex action sequences based directly on a dynamic image representation of the robot motion and the distance to a given target observation. In contrast to related works, this enables our approach to efficiently perform complex manipulation tasks from raw image observations without predefined control primitives or test time demonstrations. We report significant improvements in task success over representative MPC and IL baselines. We also demonstrate our model's generalisation capabilities in challenging, unseen tasks featuring visual noise, cluttered scenes and unseen object geometries.

We propose a method to learn 3D deformable object categories from raw single-view images, without external supervision. The method is based on an autoencoder that factors each input image into depth, albedo, viewpoint and illumination. In order to disentangle these components without supervision, we use the fact that many object categories have, at least approximately, a symmetric structure. We show that reasoning about illumination allows us to exploit the underlying object symmetry even if the appearance is not symmetric due to shading. Furthermore, we model objects that are probably, but not certainly, symmetric by predicting a symmetry probability map, learned end-to-end with the other components of the model. Our experiments show that this method can recover very accurately the 3D shape of human faces, cat faces and cars from single-view images, without any supervision or a prior shape model. Code and demo available at https://github.com/elliottwu/unsup3d.

We propose a new deep learning framework to decompose monocular videos into 3D geometry (camera pose and depth), moving objects, and their motions, with no supervision. We build upon the idea of view synthesis, which uses classical camera geometry to re-render a source image from a different point-of-view to obtain supervisory signals, specified by a predicted relative 6-degree-of-freedom pose and depth map. However, the typical view synthesis equations rely on a strong assumption: that objects in scenes do not move. This rigid-world assumption limits the predictive power, and rules out learning about objects automatically. We propose a simple solution: minimize the synthesis error on small local regions of the image instead. While the scene as a whole may be non-rigid, it is always possible to find small regions that are approximately rigid, such as inside a moving object. Our network can learn a dense pose map describing poses for each local region. This represents a significantly richer model, including 6D object motions, with little additional complexity. We establish very competitive results on unsupervised odometry and depth prediction on KITTI. We also demonstrate new capabilities on EPIC-Kitchens, a challenging dataset of indoor videos, where there is no ground truth information for depth, odometry, object segmentation or motion - yet all are recovered automatically by our approach.

Vision is one of the primary sensing modalities in autonomous driving. In this paper we look at the problem of estimating the velocity of road vehicles from a camera mounted on a moving car. Contrary to prior methods that train end-to-end deep networks that estimate the vehicles' velocity from the video pixels, we propose a two-step approach where first an off-the-shelf tracker is used to extract vehicle bounding boxes and then a small neural network is used to regress the vehicle velocity from the tracked bounding boxes. Surprisingly, we find that this still achieves state-of-the-art estimation performance with the significant benefit of separating perception from dynamics estimation via a clean, interpretable and verifiable interface which allows us distill the statistics which are crucial for velocity estimation. We show that the latter can be used to easily generate synthetic training data in the space of bounding boxes and use this to improve the performance of our method further.

The goal of self-supervised visual representation learning is to learn strong, transferable image representations, with the majority of research focusing on object or scene level. On the other hand, representation learning at part level has received significantly less attention. In this paper, we propose an unsupervised approach to object part discovery and segmentation and make three contributions. First, we construct a proxy task through a set of objectives that encourages the model to learn a meaningful decomposition of the image into its parts. Secondly, prior work argues for reconstructing or clustering pre-computed features as a proxy to parts; we show empirically that this alone is unlikely to find meaningful parts; mainly because of their low resolution and the tendency of classification networks to spatially smear out information. We suggest that image reconstruction at the level of pixels can alleviate this problem, acting as a complementary cue. Lastly, we show that the standard evaluation based on keypoint regression does not correlate well with segmentation quality and thus introduce different metrics, NMI and ARI, that better characterize the decomposition of objects into parts. Our method yields semantic parts which are consistent across fine-grained but visually distinct categories, outperforming the state of the art on three benchmark datasets. Code is available at the project page: https://www.robots.ox.ac.uk/~vgg/research/unsup-parts/.

In video transformers, the time dimension is often treated in the same way as the two spatial dimensions. However, in a scene where objects or the camera may move, a physical point imaged at one location in frame $t$ may be entirely unrelated to what is found at that location in frame $t+k$. These temporal correspondences should be modeled to facilitate learning about dynamic scenes. To this end, we propose a new drop-in block for video transformers - trajectory attention - that aggregates information along implicitly determined motion paths. We additionally propose a new method to address the quadratic dependence of computation and memory on the input size, which is particularly important for high resolution or long videos. While these ideas are useful in a range of settings, we apply them to the specific task of video action recognition with a transformer model and obtain state-of-the-art results on the Kinetics, Something-Something V2, and Epic-Kitchens datasets.

Deep learning has huge potential for accurate disease prediction with neuroimaging data, but the prediction performance is often limited by training-dataset size and computing memory requirements. To address this, we propose a deep convolutional neural network model, Simple Fully Convolutional Network (SFCN), for accurate prediction of brain age using T1-weighted structural MRI data. Compared with other popular deep network architectures, SFCN has fewer parameters, so is more compatible with small dataset size and 3D volume data. The network architecture was combined with several techniques for boosting performance, including data augmentation, pre-training, model regularization, model ensemble and prediction bias correction. We compared our overall SFCN approach with several widely-used machine learning models. It achieved state-of-the-art performance in UK Biobank data (N = 14,503), with mean absolute error (MAE) = 2.14y in brain age prediction and 99.5% in sex classification. SFCN also won (both parts of) the 2019 Predictive Analysis Challenge for brain age prediction, involving 79 competing teams (N = 2,638, MAE = 2.90y). We describe here the details of our approach, and its optimisation and validation. Our approach can easily be generalised to other tasks using different image modalities, and is released on GitHub.

We propose a new method for recognizing the pose of objects from a single image that for learning uses only unlabelled videos and a weak empirical prior on the object poses. Video frames differ primarily in the pose of the objects they contain, so our method distils the pose information by analyzing the differences between frames. The distillation uses a new dual representation of the geometry of objects as a set of 2D keypoints, and as a pictorial representation, i.e. a skeleton image. This has three benefits: (1) it provides a tight 'geometric bottleneck' which disentangles pose from appearance, (2) it can leverage powerful image-to-image translation networks to map between photometry and geometry, and (3) it allows to incorporate empirical pose priors in the learning process. The pose priors are obtained from unpaired data, such as from a different dataset or modality such as mocap, such that no annotated image is ever used in learning the pose recognition network. In standard benchmarks for pose recognition for humans and faces, our method achieves state-of-the-art performance among methods that do not require any labelled images for training. Project page: http://www.robots.ox.ac.uk/~vgg/research/unsupervised_pose/.

Saliency methods seek to explain the predictions of a model by producing an importance map across each input sample. A popular class of such methods is based on backpropagating a signal and analyzing the resulting gradient. Despite much research on such methods, relatively little work has been done to clarify the differences between such methods as well as the desiderata of these techniques. Thus, there is a need for rigorously understanding the relationships between different methods as well as their failure modes. In this work, we conduct a thorough analysis of backpropagation-based saliency methods and propose a single framework under which several such methods can be unified. As a result of our study, we make three additional contributions. First, we use our framework to propose NormGrad, a novel saliency method based on the spatial contribution of gradients of convolutional weights. Second, we combine saliency maps at different layers to test the ability of saliency methods to extract complementary information at different network levels (e.g.~trading off spatial resolution and distinctiveness) and we explain why some methods fail at specific layers (e.g., Grad-CAM anywhere besides the last convolutional layer). Third, we introduce a class-sensitivity metric and a meta-learning inspired paradigm applicable to any saliency method for improving sensitivity to the output class being explained.

Recent contributions have demonstrated that it is possible to recognize the pose of humans densely and accurately given a large dataset of poses annotated in detail. In principle, the same approach could be extended to any animal class, but the effort required for collecting new annotations for each case makes this strategy impractical, despite important applications in natural conservation, science and business. We show that, at least for proximal animal classes such as chimpanzees, it is possible to transfer the knowledge existing in dense pose recognition for humans, as well as in more general object detectors and segmenters, to the problem of dense pose recognition in other classes. We do this by (1) establishing a DensePose model for the new animal which is also geometrically aligned to humans (2) introducing a multi-head R-CNN architecture that facilitates transfer of multiple recognition tasks between classes, (3) finding which combination of known classes can be transferred most effectively to the new animal and (4) using self-calibrated uncertainty heads to generate pseudo-labels graded by quality for training a model for this class. We also introduce two benchmark datasets labelled in the manner of DensePose for the class chimpanzee and use them to evaluate our approach, showing excellent transfer learning performance.

We propose a method to learn 3D deformable object categories from raw single-view images, without external supervision. The method is based on an autoencoder that factors each input image into depth, albedo, viewpoint and illumination. In order to disentangle these components without supervision, we use the fact that many object categories have, at least in principle, a symmetric structure. We show that reasoning about illumination allows us to exploit the underlying object symmetry even if the appearance is not symmetric due to shading. Furthermore, we model objects that are probably, but not certainly, symmetric by predicting a symmetry probability map, learned end-to-end with the other components of the model. Our experiments show that this method can recover very accurately the 3D shape of human faces, cat faces and cars from single-view images, without any supervision or a prior shape model. On benchmarks, we demonstrate superior accuracy compared to another method that uses supervision at the level of 2D image correspondences.

Our goal is to collect a large-scale audio-visual dataset with low label noise from videos `in the wild' using computer vision techniques. The resulting dataset can be used for training and evaluating audio recognition models. We make three contributions. First, we propose a scalable pipeline based on computer vision techniques to create an audio dataset from open-source media. Our pipeline involves obtaining videos from YouTube; using image classification algorithms to localize audio-visual correspondence; and filtering out ambient noise using audio verification. Second, we use this pipeline to curate the VGGSound dataset consisting of more than 200k videos for 300 audio classes. Third, we investigate various Convolutional Neural Network (CNN) architectures and aggregation approaches to establish audio recognition baselines for our new dataset. Compared to existing audio datasets, VGGSound ensures audio-visual correspondence and is collected under unconstrained conditions. Code and the dataset are available at http://www.robots.ox.ac.uk/~vgg/data/vggsound/.

A large part of the current success of deep learning lies in the effectiveness of data -- more precisely: of labeled data. Yet, labelling a dataset with human annotation continues to carry high costs, especially for videos. While in the image domain, recent methods have allowed to generate meaningful (pseudo-) labels for unlabelled datasets without supervision, this development is missing for the video domain where learning feature representations is the current focus. In this work, we a) show that unsupervised labelling of a video dataset does not come for free from strong feature encoders and b) propose a novel clustering method that allows pseudo-labelling of a video dataset without any human annotations, by leveraging the natural correspondence between audio and visual modalities. An extensive analysis shows that the resulting clusters have high semantic overlap to ground truth human labels. We further introduce the first benchmarking results on unsupervised labelling of common video datasets.

We consider the problem of obtaining dense 3D reconstructions of deformable objects from single and partially occluded views. In such cases, the visual evidence is usually insufficient to identify a 3D reconstruction uniquely, so we aim at recovering several plausible reconstructions compatible with the input data. We suggest that ambiguities can be modeled more effectively by parametrizing the possible body shapes and poses via a suitable 3D model, such as SMPL for humans. We propose to learn a multi-hypothesis neural network regressor using a best-of-M loss, where each of the M hypotheses is constrained to lie on a manifold of plausible human poses by means of a generative model. We show that our method outperforms alternative approaches in ambiguous pose recovery on standard benchmarks for 3D humans, and in heavily occluded versions of these benchmarks.

We present RELATE, a model that learns to generate physically plausible scenes and videos of multiple interacting objects. Similar to other generative approaches, RELATE is trained end-to-end on raw, unlabeled data. RELATE combines an object-centric GAN formulation with a model that explicitly accounts for correlations between individual objects. This allows the model to generate realistic scenes and videos from a physically-interpretable parameterization. Furthermore, we show that modeling the object correlation is necessary to learn to disentangle object positions and identity. We find that RELATE is also amenable to physically realistic scene editing and that it significantly outperforms prior art in object-centric scene generation in both synthetic (CLEVR, ShapeStacks) and real-world data (cars). In addition, in contrast to state-of-the-art methods in object-centric generative modeling, RELATE also extends naturally to dynamic scenes and generates videos of high visual fidelity. Source code, datasets and more results are available at http://geometry.cs.ucl.ac.uk/projects/2020/relate/.

The increasing impact of black box models, and particularly of unsupervised ones, comes with an increasing interest in tools to understand and interpret them. In this paper, we consider in particular how to characterise visual groupings discovered automatically by deep neural networks, starting with state-of-the-art clustering methods. In some cases, clusters readily correspond to an existing labelled dataset. However, often they do not, yet they still maintain an "intuitive interpretability''. We introduce two concepts, visual learnability and describability, that can be used to quantify the interpretability of arbitrary image groupings, including unsupervised ones. The idea is to measure (1) how well humans can learn to reproduce a grouping by measuring their ability to generalise from a small set of visual examples (learnability) and (2) whether the set of visual examples can be replaced by a succinct, textual description (describability). By assessing human annotators as classifiers, we remove the subjective quality of existing evaluation metrics. For better scalability, we finally propose a class-level captioning system to generate descriptions for visual groupings automatically and compare it to human annotators using the describability metric.

In this work, we focus on the task of learning and representing dense correspondences in deformable object categories. While this problem has been considered before, solutions so far have been rather ad-hoc for specific object types (i.e., humans), often with significant manual work involved. However, scaling the geometry understanding to all objects in nature requires more automated approaches that can also express correspondences between related, but geometrically different objects. To this end, we propose a new, learnable image-based representation of dense correspondences. Our model predicts, for each pixel in a 2D image, an embedding vector of the corresponding vertex in the object mesh, therefore establishing dense correspondences between image pixels and 3D object geometry. We demonstrate that the proposed approach performs on par or better than the state-of-the-art methods for dense pose estimation for humans, while being conceptually simpler. We also collect a new in-the-wild dataset of dense correspondences for animal classes and demonstrate that our framework scales naturally to the new deformable object categories.

We propose the Canonical 3D Deformer Map, a new representation of the 3D shape of common object categories that can be learned from a collection of 2D images of independent objects. Our method builds in a novel way on concepts from parametric deformation models, non-parametric 3D reconstruction, and canonical embeddings, combining their individual advantages. In particular, it learns to associate each image pixel with a deformation model of the corresponding 3D object point which is canonical, i.e. intrinsic to the identity of the point and shared across objects of the category. The result is a method that, given only sparse 2D supervision at training time, can, at test time, reconstruct the 3D shape and texture of objects from single views, while establishing meaningful dense correspondences between object instances. It also achieves state-of-the-art results in dense 3D reconstruction on public in-the-wild datasets of faces, cars, and birds.

In this paper, a novel benchmark is introduced for evaluating local image descriptors. We demonstrate limitations of the commonly used datasets and evaluation protocols, that lead to ambiguities and contradictory results in the literature. Furthermore, these benchmarks are nearly saturated due to the recent improvements in local descriptors obtained by learning from large annotated datasets. To address these issues, we introduce a new large dataset suitable for training and testing modern descriptors, together with strictly defined evaluation protocols in several tasks such as matching, retrieval and verification. This allows for more realistic, thus more reliable comparisons in different application scenarios. We evaluate the performance of several state-of-the-art descriptors and analyse their properties. We show that a simple normalisation of traditional hand-crafted descriptors is able to boost their performance to the level of deep learning based descriptors once realistic benchmarks are considered. Additionally we specify a protocol for learning and evaluating using cross validation. We show that when training state-of-the-art descriptors on this dataset, the traditional verification task is almost entirely saturated.

We propose an unsupervised method to learn the 3D geometry of object categories by looking around them. Differently from traditional approaches, this method does not require CAD models or manual supervision. Instead, using only video sequences showing object instances from a moving viewpoint, the method learns a deep neural network that can predict several aspects of the 3D geometry of such objects from single images. The network has three components. The first is a Siamese viewpoint factorization network that robustly aligns the input videos and learns to predict the absolute viewpoint of the object from a single image. The second is a depth estimation network that performs monocular depth prediction. The third is a shape completion network that predicts the full 3D shape of the object from the output of the monocular depth prediction module. While the three modules solve very different task, we show that they all benefit significantly from allowing networks to perform probabilistic predictions. This results in a self-assessment mechanism which is crucial for obtaining high quality predictions. Our network achieves state-of-the-art results on viewpoint prediction, depth estimation, and 3D point cloud estimation on public benchmarks.

DensePose supersedes traditional landmark detectors by densely mapping image pixels to body surface coordinates. This power, however, comes at a greatly increased annotation cost, as supervising the model requires to manually label hundreds of points per pose instance. In this work, we thus seek methods to significantly slim down the DensePose annotations, proposing more efficient data collection strategies. In particular, we demonstrate that if annotations are collected in video frames, their efficacy can be multiplied for free by using motion cues. To explore this idea, we introduce DensePose-Track, a dataset of videos where selected frames are annotated in the traditional DensePose manner. Then, building on geometric properties of the DensePose mapping, we use the video dynamic to propagate ground-truth annotations in time as well as to learn from Siamese equivariance constraints. Having performed exhaustive empirical evaluation of various data annotation and learning strategies, we demonstrate that doing so can deliver significantly improved pose estimation results over strong baselines. However, despite what is suggested by some recent works, we show that merely synthesizing motion patterns by applying geometric transformations to isolated frames is significantly less effective, and that motion cues help much more when they are extracted from videos.

Attribution is the problem of finding which parts of an image are the most responsible for the output of a deep neural network. An important family of attribution methods is based on measuring the effect of perturbations applied to the input image, either via exhaustive search or by finding representative perturbations via optimization. In this paper, we discuss some of the shortcomings of existing approaches to perturbation analysis and address them by introducing the concept of extremal perturbations, which are theoretically grounded and interpretable. We also introduce a number of technical innovations to compute these extremal perturbations, including a new area constraint and a parametric family of smooth perturbations, which allow us to remove all tunable weighing factors from the optimization problem. We analyze the effect of perturbations as a function of their area, demonstrating excellent sensitivity to the spatial properties of the network under stimulation. We also extend perturbation analysis to the intermediate layers of a deep neural network. This application allows us to show how compactly an image can be represented (in terms of the number of channels it requires). We also demonstrate that the consistency with which images of a given class rely on the same intermediate channel correlates well with class accuracy.

We consider the problem of discovering novel object categories in an image collection. While these images are unlabelled, we also assume prior knowledge of related but different image classes. We use such prior knowledge to reduce the ambiguity of clustering, and improve the quality of the newly discovered classes. Our contributions are twofold. The first contribution is to extend Deep Embedded Clustering to a transfer learning setting; we also improve the algorithm by introducing a representation bottleneck, temporal ensembling, and consistency. The second contribution is a method to estimate the number of classes in the unlabelled data. This also transfers knowledge from the known classes, using them as probes to diagnose different choices for the number of classes in the unlabelled subset. We thoroughly evaluate our method, substantially outperforming state-of-the-art techniques in a large number of benchmarks, including ImageNet, OmniGlot, CIFAR-100, CIFAR-10, and SVHN.

We propose a fast second-order method that can be used as a drop-in replacement for current deep learning solvers. Compared to stochastic gradient descent (SGD), it only requires two additional forward-mode automatic differentiation operations per iteration, which has a computational cost comparable to two standard forward passes and is easy to implement. Our method addresses long-standing issues with current second-order solvers, which invert an approximate Hessian matrix every iteration exactly or by conjugate-gradient methods, procedures that are much slower than a SGD step. Instead, we propose to keep a single estimate of the gradient projected by the inverse Hessian matrix, and update it once per iteration with just two passes over the network. This estimate has the same size and is similar to the momentum variable that is commonly used in SGD. No estimate of the Hessian is maintained. We first validate our method, called CURVEBALL, on small problems with known solutions (noisy Rosenbrock function and degenerate 2-layer linear networks), where current deep learning solvers struggle. We then train several large models on CIFAR and ImageNet, including ResNet and VGG-f networks, where we demonstrate faster convergence with no hyperparameter tuning. We also show our optimiser's generality by testing on a large set of randomly-generated architectures.

We present a novel clustering objective that learns a neural network classifier from scratch, given only unlabelled data samples. The model discovers clusters that accurately match semantic classes, achieving state-of-the-art results in eight unsupervised clustering benchmarks spanning image classification and segmentation. These include STL10, an unsupervised variant of ImageNet, and CIFAR10, where we significantly beat the accuracy of our closest competitors by 6.6 and 9.5 absolute percentage points respectively. The method is not specialised to computer vision and operates on any paired dataset samples; in our experiments we use random transforms to obtain a pair from each image. The trained network directly outputs semantic labels, rather than high dimensional representations that need external processing to be usable for semantic clustering. The objective is simply to maximise mutual information between the class assignments of each pair. It is easy to implement and rigorously grounded in information theory, meaning we effortlessly avoid degenerate solutions that other clustering methods are susceptible to. In addition to the fully unsupervised mode, we also test two semi-supervised settings. The first achieves 88.8% accuracy on STL10 classification, setting a new global state-of-the-art over all existing methods (whether supervised, semi-supervised or unsupervised). The second shows robustness to 90% reductions in label coverage, of relevance to applications that wish to make use of small amounts of labels. github.com/xu-ji/IIC.

We propose C3DPO, a method for extracting 3D models of deformable objects from 2D keypoint annotations in unconstrained images. We do so by learning a deep network that reconstructs a 3D object from a single view at a time, accounting for partial occlusions, and explicitly factoring the effects of viewpoint changes and object deformations. In order to achieve this factorization, we introduce a novel regularization technique. We first show that the factorization is successful if, and only if, there exists a certain canonicalization function of the reconstructed shapes. Then, we learn the canonicalization function together with the reconstruction one, which constrains the result to be consistent. We demonstrate state-of-the-art reconstruction results for methods that do not use ground-truth 3D supervision for a number of benchmarks, including Up3D and PASCAL3D+.

Equivariance to random image transformations is an effective method to learn landmarks of object categories, such as the eyes and the nose in faces, without manual supervision. However, this method does not explicitly guarantee that the learned landmarks are consistent with changes between different instances of the same object, such as different facial identities. In this paper, we develop a new perspective on the equivariance approach by noting that dense landmark detectors can be interpreted as local image descriptors equipped with invariance to intra-category variations. We then propose a direct method to enforce such an invariance in the standard equivariant loss. We do so by exchanging descriptor vectors between images of different object instances prior to matching them geometrically. In this manner, the same vectors must work regardless of the specific object identity considered. We use this approach to learn vectors that can simultaneously be interpreted as local descriptors and dense landmarks, combining the advantages of both. Experiments on standard benchmarks show that this approach can match, and in some cases surpass state-of-the-art performance amongst existing methods that learn landmarks without supervision. Code is available at www.robots.ox.ac.uk/~vgg/research/DVE/.

Many machine learning methods depend on human supervision to achieve optimal performance. However, in tasks such as DensePose, where the goal is to establish dense visual correspondences between images, the quality of manual annotations is intrinsically limited. We address this issue by augmenting neural network predictors with the ability to output a distribution over labels, thus explicitly and introspectively capturing the aleatoric uncertainty in the annotations. Compared to previous works, we show that correlated error fields arise naturally in applications such as DensePose and these fields can be modeled by deep networks, leading to a better understanding of the annotation errors. We show that these models, by understanding uncertainty better, can solve the original DensePose task more accurately, thus setting the new state-of-the-art accuracy in this benchmark. Finally, we demonstrate the utility of the uncertainty estimates in fusing the predictions of produced by multiple models, resulting in a better and more principled approach to model ensembling which can further improve accuracy.

Data-augmentation is key to the training of neural networks for image classification. This paper first shows that existing augmentations induce a significant discrepancy between the size of the objects seen by the classifier at train and test time: in fact, a lower train resolution improves the classification at test time! We then propose a simple strategy to optimize the classifier performance, that employs different train and test resolutions. It relies on a computationally cheap fine-tuning of the network at the test resolution. This enables training strong classifiers using small training images, and therefore significantly reduce the training time. For instance, we obtain 77.1% top-1 accuracy on ImageNet with a ResNet-50 trained on 128x128 images, and 79.8% with one trained at 224x224. A ResNeXt-101 32x48d pre-trained with weak supervision on 940 million 224x224 images and further optimized with our technique for test resolution 320x320 achieves 86.4% top-1 accuracy (top-5: 98.0%). To the best of our knowledge this is the highest ImageNet single-crop accuracy to date.

While the basic laws of Newtonian mechanics are well understood, explaining a physical scenario still requires manually modeling the problem with suitable equations and estimating the associated parameters. In order to be able to leverage the approximation capabilities of artificial intelligence techniques in such physics related contexts, researchers have handcrafted relevant states, and then used neural networks to learn the state transitions using simulation runs as training data. Unfortunately, such approaches are unsuited for modeling complex real-world scenarios, where manually authoring relevant state spaces tend to be tedious and challenging. In this work, we investigate if neural networks can implicitly learn physical states of real-world mechanical processes only based on visual data while internally modeling non-homogeneous environment and in the process enable long-term physical extrapolation. We develop a recurrent neural network architecture for this task and also characterize resultant uncertainties in the form of evolving variance estimates. We evaluate our setup, to extrapolate motion of rolling ball(s) on bowls of varying shape and orientation, and on arbitrary heightfields using only images as input. We report significant improvements over existing image-based methods both in terms of accuracy of predictions and complexity of scenarios; and report competitive performance with approaches that, unlike us, assume access to internal physical states.

Despite the importance of image representations such as histograms of oriented gradients and deep Convolutional Neural Networks (CNN), our theoretical understanding of them remains limited. Aimed at filling this gap, we investigate two key mathematical properties of representations: equivariance and equivalence. Equivariance studies how transformations of the input image are encoded by the representation, invariance being a special case where a transformation has no effect. Equivalence studies whether two representations, for example two different parameterizations of a CNN, two different layers, or two different CNN architectures, share the same visual information or not. A number of methods to establish these properties empirically are proposed, including introducing transformation and stitching layers in CNNs. These methods are then applied to popular representations to reveal insightful aspects of their structure, including clarifying at which layers in a CNN certain geometric invariances are achieved and how various CNN architectures differ. We identify several predictors of geometric and architectural compatibility, including the spatial resolution of the representation and the complexity and depth of the models. While the focus of the paper is theoretical, direct applications to structured-output regression are demonstrated too.

We present a new autoencoder-type architecture that is trainable in an unsupervised mode, sustains both generation and inference, and has the quality of conditional and unconditional samples boosted by adversarial learning. Unlike previous hybrids of autoencoders and adversarial networks, the adversarial game in our approach is set up directly between the encoder and the generator, and no external mappings are trained in the process of learning.The game objective compares the divergences of each of the real and the generated data distributions with the prior distribution in the latent space. We show that direct generator-vs-encoder game leads to a tight coupling of the two components, resulting in samples and reconstructions of a comparable quality to some recently-proposed more complex architectures.

In an effort to understand the meaning of the intermediate representations captured by deep networks, recent papers have tried to associate specific semantic concepts to individual neural network filter responses, where interesting correlations are often found, largely by focusing on extremal filter responses. In this paper, we show that this approach can favor easy-to-interpret cases that are not necessarily representative of the average behavior of a representation. A more realistic but harder-to-study hypothesis is that semantic representations are distributed, and thus filters must be studied in conjunction. In order to investigate this idea while enabling systematic visualization and quantification of multiple filter responses, we introduce the Net2Vec framework, in which semantic concepts are mapped to vectorial embeddings based on corresponding filter responses. By studying such embeddings, we are able to show that 1., in most cases, multiple filters are required to code for a concept, that 2., often filters are not concept specific and help encode multiple concepts, and that 3., compared to single filter activations, filter embeddings are able to better characterize the meaning of a representation and its relationship to other concepts.

Autonomous agents need to reason about the world beyond their instantaneous sensory input. Integrating information over time, however, requires switching from an egocentric representation of a scene to an allocentric one, expressed in the world reference frame. It must also be possible to update the representation dynamically, which requires localizing and registering the sensor with respect to the world reference. In this paper, we develop a differentiable module that satisfies such requirements, while being robust, efficient, and suitable for integration in end-to-end deep networks. The module contains an allocentric spatial memory that can be accessed associatively by feeding to it the current sensory input, resulting in localization, and then updated using an LSTM or similar mechanism. We formulate efficient localization and registration of sensory information as a dual pair of convolution/deconvolution operators in memory space. The map itself is a 2.5D representation of an environment storing information that a deep neural network module learns to distill from RGBD input. The result is a map that contains multi-task information, different from classical approaches to mapping such as structure-from-motion. We present results using synthetic mazes, a dataset of hours of recorded gameplay of the classic game Doom, and the very recent Active Vision Dataset of real images captured from a robot.

Self-supervision can dramatically cut back the amount of manually-labelled data required to train deep neural networks. While self-supervision has usually been considered for tasks such as image classification, in this paper we aim at extending it to geometry-oriented tasks such as semantic matching and part detection. We do so by building on several recent ideas in unsupervised landmark detection. Our approach learns dense distinctive visual descriptors from an unlabeled dataset of images using synthetic image transformations. It does so by means of a robust probabilistic formulation that can introspectively determine which image regions are likely to result in stable image matching. We show empirically that a network pre-trained in this manner requires significantly less supervision to learn semantic object parts compared to numerous pre-training alternatives. We also show that the pre-trained representation is excellent for semantic object matching.

A practical limitation of deep neural networks is their high degree of specialization to a single task and visual domain. Recently, inspired by the successes of transfer learning, several authors have proposed to learn instead universal feature extractors that, used as the first stage of any deep network, work well for several tasks and domains simultaneously. Nevertheless, such universal features are still somewhat inferior to specialized networks. To overcome this limitation, in this paper we propose to consider instead universal parametric families of neural networks, which still contain specialized problem-specific models, but differing only by a small number of parameters. We study different designs for such parametrizations, including series and parallel residual adapters, joint adapter compression, and parameter allocations, and empirically identify the ones that yield the highest compression. We show that, in order to maximize performance, it is necessary to adapt both shallow and deep layers of a deep network, but the required changes are very small. We also show that these universal parametrization are very effective for transfer learning, where they outperform traditional fine-tuning techniques.

Deep convolutional networks have become a popular tool for image generation and restoration. Generally, their excellent performance is imputed to their ability to learn realistic image priors from a large number of example images. In this paper, we show that, on the contrary, the structure of a generator network is sufficient to capture a great deal of low-level image statistics prior to any learning. In order to do so, we show that a randomly-initialized neural network can be used as a handcrafted prior with excellent results in standard inverse problems such as denoising, superresolution, and inpainting. Furthermore, the same prior can be used to invert deep neural representations to diagnose them, and to restore images based on flash-no flash input pairs. Apart from its diverse applications, our approach highlights the inductive bias captured by standard generator network architectures. It also bridges the gap between two very popular families of image restoration methods: learning-based methods using deep convolutional networks and learning-free methods based on handcrafted image priors such as self-similarity.

Obtaining large, human labelled speech datasets to train models for emotion recognition is a notoriously challenging task, hindered by annotation cost and label ambiguity. In this work, we consider the task of learning embeddings for speech classification without access to any form of labelled audio. We base our approach on a simple hypothesis: that the emotional content of speech correlates with the facial expression of the speaker. By exploiting this relationship, we show that annotations of expression can be transferred from the visual domain (faces) to the speech domain (voices) through cross-modal distillation. We make the following contributions: (i) we develop a strong teacher network for facial emotion recognition that achieves the state of the art on a standard benchmark; (ii) we use the teacher to train a student, tabula rasa, to learn representations (embeddings) for speech emotion recognition without access to labelled audio data; and (iii) we show that the speech emotion embedding can be used for speech emotion recognition on external benchmark datasets. Code, models and data are available.

While the use of bottom-up local operators in convolutional neural networks (CNNs) matches well some of the statistics of natural images, it may also prevent such models from capturing contextual long-range feature interactions. In this work, we propose a simple, lightweight approach for better context exploitation in CNNs. We do so by introducing a pair of operators: gather, which efficiently aggregates feature responses from a large spatial extent, and excite, which redistributes the pooled information to local features. The operators are cheap, both in terms of number of added parameters and computational complexity, and can be integrated directly in existing architectures to improve their performance. Experiments on several datasets show that gather-excite can bring benefits comparable to increasing the depth of a CNN at a fraction of the cost. For example, we find ResNet-50 with gather-excite operators is able to outperform its 101-layer counterpart on ImageNet with no additional learnable parameters. We also propose a parametric gather-excite operator pair which yields further performance gains, relate it to the recently-introduced Squeeze-and-Excitation Networks, and analyse the effects of these changes to the CNN feature activation statistics.

We propose a method for learning landmark detectors for visual objects (such as the eyes and the nose in a face) without any manual supervision. We cast this as the problem of generating images that combine the appearance of the object as seen in a first example image with the geometry of the object as seen in a second example image, where the two examples differ by a viewpoint change and/or an object deformation. In order to factorize appearance and geometry, we introduce a tight bottleneck in the geometry-extraction process that selects and distils geometry-related features. Compared to standard image generation problems, which often use generative adversarial networks, our generation task is conditioned on both appearance and geometry and thus is significantly less ambiguous, to the point that adopting a simple perceptual loss formulation is sufficient. We demonstrate that our approach can learn object landmarks from synthetic image deformations or videos, all without manual supervision, while outperforming state-of-the-art unsupervised landmark detectors. We further show that our method is applicable to a large variety of datasets - faces, people, 3D objects, and digits - without any modifications.

We propose a new approach to model and learn, without manual supervision, the symmetries of natural objects, such as faces or flowers, given only images as input. It is well known that objects that have a symmetric structure do not usually result in symmetric images due to articulation and perspective effects. This is often tackled by seeking the intrinsic symmetries of the underlying 3D shape, which is very difficult to do when the latter cannot be recovered reliably from data. We show that, if only raw images are given, it is possible to look instead for symmetries in the space of object deformations. We can then learn symmetries from an unstructured collection of images of the object as an extension of the recently-introduced object frame representation, modified so that object symmetries reduce to the obvious symmetry groups in the normalized space. We also show that our formulation provides an explanation of the ambiguities that arise in recovering the pose of symmetric objects from their shape or images and we provide a way of discounting such ambiguities in learning.

We introduce the concept of dynamic image, a novel compact representation of videos useful for video analysis, particularly in combination with convolutional neural networks (CNNs). A dynamic image encodes temporal data such as RGB or optical flow videos by using the concept of 'rank pooling'. The idea is to learn a ranking machine that captures the temporal evolution of the data and to use the parameters of the latter as a representation. We call the resulting representation dynamic image because it summarizes the video dynamics in addition to appearance. This powerful idea allows to convert any video to an image so that existing CNN models pre-trained with still images can be immediately extended to videos. We also present an efficient approximate rank pooling operator that runs two orders of magnitude faster than the standard ones with any loss in ranking performance and can be formulated as a CNN layer. To demonstrate the power of the representation, we introduce a novel four stream CNN architecture which can learn from RGB and optical flow frames as well as from their dynamic image representations. We show that the proposed network achieves state-of-the-art performance, 95.5 and 72.5 percent accuracy, in the UCF101 and HMDB51, respectively.

In this paper, a novel benchmark is introduced for evaluating local image descriptors. We demonstrate limitations of the commonly used datasets and evaluation protocols, that lead to ambiguities and contradictory results in the literature. Furthermore, these benchmarks are nearly saturated due to the recent improvements in local descriptors obtained by learning from large annotated datasets. To address these issues, we introduce a new large dataset suitable for training and testing modern descriptors, together with strictly defined evaluation protocols in several tasks such as matching, retrieval and verification. This allows for more realistic, thus more reliable comparisons in different application scenarios. We evaluate the performance of several state-of-the-art descriptors and analyse their properties. We show that a simple normalisation of traditional hand-crafted descriptors is able to boost their performance to the level of deep learning based descriptors once realistic benchmarks are considered. Additionally we specify a protocol for learning and evaluating using cross validation. We show that when training state-of-the-art descriptors on this dataset, the traditional verification task is almost entirely saturated.

Despite significant progress of deep learning in recent years, state-of-the-art semantic matching methods still rely on legacy features such as SIFT or HoG. We argue that the strong invariance properties that are key to the success of recent deep architectures on the classification task make them unfit for dense correspondence tasks, unless a large amount of supervision is used. In this work, we propose a deep network, termed AnchorNet, that produces image representations that are well-suited for semantic matching. It relies on a set of filters whose response is geometrically consistent across different object instances, even in the presence of strong intra-class, scale, or viewpoint variations. Trained only with weak image-level labels, the final representation successfully captures information about the object structure and improves results of state-of-the-art semantic matching methods such as the Deformable Spatial Pyramid or the Proposal Flow methods. We show positive results on the cross-instance matching task where different instances of the same object category are matched as well as on a new cross-category semantic matching task aligning pairs of instances each from a different object class.

The Correlation Filter is an algorithm that trains a linear template to discriminate between images and their translations. It is well suited to object tracking because its formulation in the Fourier domain provides a fast solution, enabling the detector to be re-trained once per frame. Previous works that use the Correlation Filter, however, have adopted features that were either manually designed or trained for a different task. This work is the first to overcome this limitation by interpreting the Correlation Filter learner, which has a closed-form solution, as a differentiable layer in a deep neural network. This enables learning deep features that are tightly coupled to the Correlation Filter. Experiments illustrate that our method has the important practical benefit of allowing lightweight architectures to achieve state-of-the-art performance at high framerates.

As machine learning algorithms are increasingly applied to high impact yet high risk tasks, such as medical diagnosis or autonomous driving, it is critical that researchers can explain how such algorithms arrived at their predictions. In recent years, a number of image saliency methods have been developed to summarize where highly complex neural networks “look” in an image for evidence for their predictions. However, these techniques are limited by their heuristic nature and architectural constraints. In this paper, we make two main contributions: First, we propose a general framework for learning different kinds of explanations for any black box algorithm. Second, we specialise the framework to find the part of an image most responsible for a classifier decision. Unlike previous works, our method is model-agnostic and testable because it is grounded in explicit and interpretable image perturbations.

Traditional approaches for learning 3D object categories use either synthetic data or manual supervision. In this paper, we propose a method which does not require manual annotations and is instead cued by observing objects from a moving vantage point. Our system builds on two innovations: a Siamese viewpoint factorization network that robustly aligns different videos together without explicitly comparing 3D shapes; and a 3D shape completion network that can extract the full shape of an object from partial observations. We also demonstrate the benefits of configuring networks to perform probabilistic predictions as well as of geometry-aware data augmentation schemes. We obtain state-of-the-art results on publicly-available benchmarks.

Learning automatically the structure of object categories remains an important open problem in computer vision. In this paper, we propose a novel unsupervised approach that can discover and learn landmarks in object categories, thus characterizing their structure. Our approach is based on factorizing image deformations, as induced by a viewpoint change or an object deformation, by learning a deep neural network that detects landmarks consistently with such visual effects. Furthermore, we show that the learned landmarks establish meaningful correspondences between different object instances in a category without having to impose this requirement explicitly. We assess the method qualitatively on a variety of object types, natural and man-made. We also show that our unsupervised landmarks are highly predictive of manually-annotated landmarks in face benchmark datasets, and can be used to regress these with a high degree of accuracy.

Convolutional Neural Networks (CNNs) are extremely efficient, since they exploit the inherent translation-invariance of natural images. However, translation is just one of a myriad of useful spatial transformations. Can the same efficiency be attained when considering other spatial invariances? Such generalized convolutions have been considered in the past, but at a high computational cost. We present a construction that is simple and exact, yet has the same computational complexity that standard convolutions enjoy. It consists of a constant image warp followed by a simple convolution, which are standard blocks in deep learning toolboxes. With a carefully crafted warp, the resulting architecture can be made equivariant to a wide range of two-parameter spatial transformations. We show encouraging results in realistic scenarios, including the estimation of vehicle poses in the Google Earth dataset (rotation and scale), and face poses in Annotated Facial Landmarks in the Wild (3D rotations under perspective).

There is a growing interest in learning data representations that work well for many different types of problems and data. In this paper, we look in particular at the task of learning a single visual representation that can be successfully utilized in the analysis of very different types of images, from dog breeds to stop signs and digits. Inspired by recent work on learning networks that predict the parameters of another, we develop a tunable deep network architecture that, by means of adapter residual modules, can be steered on the fly to diverse visual domains. Our method achieves a high degree of parameter sharing while maintaining or even improving the accuracy of domain-specific representations. We also introduce the Visual Decathlon Challenge, a benchmark that evaluates the ability of representations to capture simultaneously ten very different visual domains and measures their ability to recognize well uniformly.

One of the key challenges of visual perception is to extract abstract models of 3D objects and object categories from visual measurements, which are affected by complex nuisance factors such as viewpoint, occlusion, motion, and deformations. Starting from the recent idea of viewpoint factorization, we propose a new approach that, given a large number of images of an object and no other supervision, can extract a dense object-centric coordinate frame. This coordinate frame is invariant to deformations of the images and comes with a dense equivariant labelling neural network that can map image pixels to their corresponding object coordinates. We demonstrate the applicability of this method to simple articulated objects and deformable objects such as human faces, learning embeddings from random synthetic transformations or optical flow correspondences, all without any manual supervision.

We introduce the concept of dynamic image, a novel compact representation of videos useful for video analysis especially when convolutional neural networks (CNNs) are used. The dynamic image is based on the rank pooling concept and is obtained through the parameters of a ranking machine that encodes the temporal evolution of the frames of the video. Dynamic images are obtained by directly applying rank pooling on the raw image pixels of a video producing a single RGB image per video. This idea is simple but powerful as it enables the use of existing CNN models directly on video data with fine-tuning. We present an efficient and effective approximate rank pooling operator, speeding it up orders of magnitude compared to rank pooling. Our new approximate rank pooling CNN layer allows us to generalize dynamic images to dynamic feature maps and we demonstrate the power of our new representations on standard benchmarks in action recognition achieving state-of-the-art performance.

Weakly supervised learning of object detection is an important problem in image understanding that still does not have a satisfactory solution. In this paper, we address this problem by exploiting the power of deep convolutional neural networks pre-trained on large-scale image-level classification tasks. We propose a weakly supervised deep detection architecture that modifies one such network to operate at the level of image regions, performing simultaneously region selection and classification. Trained as an image classifier, the architecture implicitly learns object detectors that are better than alternative weakly supervised detection systems on the PASCAL VOC data. The model, which is a simple and elegant end-to-end architecture, outperforms standard data augmentation and fine-tuning techniques for the task of image-level classification as well.

In this paper we introduce a new method for text detection in natural images. The method comprises two contributions: First, a fast and scalable engine to generate synthetic images of text in clutter. This engine overlays synthetic text to existing background images in a natural way, accounting for the local 3D scene geometry. Second, we use the synthetic images to train a Fully-Convolutional Regression Network (FCRN) which efficiently performs text detection and bounding-box regression at all locations and multiple scales in an image. We discuss the relation of FCRN to the recently-introduced YOLO detector, as well as other end-to-end object detection systems based on deep learning. The resulting detection network significantly out performs current methods for text detection in natural images, achieving an F-measure of 84.2% on the standard ICDAR 2013 benchmark. Furthermore, it can process 15 images per second on a GPU.

Gatys et al. recently demonstrated that deep networks can generate beautiful textures and stylized images from a single texture example. However, their methods requires a slow and memory-consuming optimization process. We propose here an alternative approach that moves the computational burden to a learning stage. Given a single example of a texture, our approach trains compact feed-forward convolutional networks to generate multiple samples of the same texture of arbitrary size and to transfer artistic style from a given image to any other image. The resulting networks are remarkably light-weight and can generate textures of quality comparable to Gatys et al., but hundreds of times faster. More generally, our approach highlights the power and flexibility of generative feed-forward models trained with complex and expressive loss functions.

One-shot learning is usually tackled by using generative models or discriminative embeddings. Discriminative methods based on deep learning, which are very effective in other learning scenarios, are ill-suited for one-shot learning as they need large amounts of training data. In this paper, we propose a method to learn the parameters of a deep model in one shot. We construct the learner as a second deep network, called a learnet, which predicts the parameters of a pupil network from a single exemplar. In this manner we obtain an efficient feed-forward one-shot learner, trained end-to-end by minimizing a one-shot classification objective in a learning to learn formulation. In order to make the construction feasible, we propose a number of factorizations of the parameters of the pupil network. We demonstrate encouraging results by learning characters from single exemplars in Omniglot, and by tracking visual objects from a single initial exemplar in the Visual Object Tracking benchmark.

Modern discriminative predictors have been shown to match natural intelligences in specific perceptual tasks in image classification, object and part detection, boundary extraction, etc. However, a major advantage that natural intelligences still have is that they work well for all perceptual problems together, solving them efficiently and coherently in an integrated manner. In order to capture some of these advantages in machine perception, we ask two questions: whether deep neural networks can learn universal image representations, useful not only for a single task but for all of them, and how the solutions to the different tasks can be integrated in this framework. We answer by proposing a new architecture, which we call multinet, in which not only deep image features are shared between tasks, but where tasks can interact in a recurrent manner by encoding the results of their analysis in a common shared representation of the data. In this manner, we show that the performance of individual tasks in standard benchmarks can be improved first by sharing features between them and then, more significantly, by integrating their solutions in the common representation.

Visual textures have played a key role in image understanding because they convey important semantics of images, and because texture representations that pool local image descriptors in an orderless manner have had a tremendous impact in diverse applications. In this paper we make several contributions to texture understanding. First, instead of focusing on texture instance and material category recognition, we propose a human-interpretable vocabulary of texture attributes to describe common texture patterns, complemented by a new describable texture dataset for benchmarking. Second, we look at the problem of recognizing materials and texture attributes in realistic imaging conditions, including when textures appear in clutter, developing corresponding benchmarks on top of the recently proposed OpenSurfaces dataset. Third, we revisit classic texture represenations, including bag-of-visual-words and the Fisher vectors, in the context of deep learning and show that these have excellent efficiency and generalization properties if the convolutional layers of a deep model are used as filter banks. We obtain in this manner state-of-the-art performance in numerous datasets well beyond textures, an efficient method to apply deep features to image regions, as well as benefit in transferring features from one domain to another.

Deep convolutional neural networks (CNNs) have had a major impact in most areas of image understanding, including object category detection. In object detection, methods such as R-CNN have obtained excellent results by integrating CNNs with region proposal generation algorithms such as selective search. In this paper, we investigate the role of proposal generation in CNN-based detectors in order to determine whether it is a necessary modelling component, carrying essential geometric information not contained in the CNN, or whether it is merely a way of accelerating detection. We do so by designing and evaluating a detector that uses a trivial region generation scheme, constant for each image. Combined with SPP, this results in an excellent and fast detector that does not require to process an image with algorithms other than the CNN itself. We also streamline and simplify the training of CNN-based detectors by integrating several learning steps in a single algorithm, as well as by proposing a number of improvements that accelerate detection.

The goal of this paper is face recognition – from either a single photograph or from a set of faces tracked in a video. Recent progress in this area has been due to two factors: (i) end to end learning for the task using a convolutional neural network (CNN), and (ii) the availability of very large scale training datasets. We make two contributions: first, we show how a very large scale dataset (2.6M images, over 2.6K people) can be assembled by a combination of automation and human in the loop, and discuss the trade off between data purity and time; second, we traverse through the complexities of deep network training and face recognition to present methods and procedures to achieve comparable state of the art results on the standard LFW and YTF face benchmarks.

Research in texture recognition often concentrates on the problem of material recognition in uncluttered conditions, an assumption rarely met by applications. In this work we conduct a first study of material and describable texture attributes recognition in clutter, using a new dataset derived from the OpenSurface texture repository. Motivated by the challenge posed by this problem, we propose a new texture descriptor, FV-CNN, obtained by Fisher Vector pooling of a Convolutional Neural Network (CNN) filter bank. FV-CNN substantially improves the state-of-the-art in texture, material and scene recognition. Our approach achieves 79.8% accuracy on Flickr material dataset and 81% accuracy on MIT indoor scenes, providing absolute gains of more than 10% over existing approaches. FV-CNN easily transfers across domains without requiring feature adaptation as for methods that build on the fully-connected layers of CNNs. Furthermore, FV-CNN can seamlessly incorporate multi-scale information and describe regions of arbitrary shapes and sizes. Our approach is particularly suited at localizing “stuff” categories and obtains state-of-the-art results on MSRC segmentation dataset, as well as promising results on recognizing materials and surface attributes in clutter on the OpenSurfaces dataset.

Despite the importance of image representations such as histograms of oriented gradients and deep Convolutional Neural Networks (CNN), our theoretical understanding of them remains limited. Aiming at filling this gap, we investigate three key mathematical properties of representations: equivariance, invariance, and equivalence. Equivariance studies how transformations of the input image are encoded by the representation, invariance being a special case where a transformation has no effect. Equivalence studies whether two representations, for example two different parametrisations of a CNN, capture the same visual information or not. A number of methods to establish these properties empirically are proposed, including introducing transformation and stitching layers in CNNs. These methods are then applied to popular representations to reveal insightful aspects of their structure, including clarifying at which layers in a CNN certain geometric invariances are achieved. While the focus of the paper is theoretical, direct applications to structured-output regression are demonstrated too.

Image representations, from SIFT and Bag of Visual Words to Convolutional Neural Networks (CNNs), are a crucial component of almost any image understanding system. Nevertheless, our understanding of them remains limited. In this paper we conduct a direct analysis of the visual information contained in representations by asking the following question: given an encoding of an image, to which extent is it possible to reconstruct the image itself? To answer this question we contribute a general framework to invert representations. We show that this method can invert representations such as HOG more accurately than recent alternatives while being applicable to CNNs too. We then use this technique to study the inverse of recent state-of-the-art CNN image representations for the first time. Among our findings, we show that several layers in CNNs retain photographically accurate information about the image, with different degrees of geometric and photometric invariance.

MatConvNet is an open source implementation of Convolutional Neural Networks (CNNs) with a deep integration in the MATLAB environment. The toolbox is designed with an emphasis on simplicity and flexibility. It exposes the building blocks of CNNs as easy-to-use MATLAB functions, providing routines for computing convolutions with filter banks, feature pooling, normalisation, and much more. MatConvNet can be easily extended, often using only MATLAB code, allowing fast prototyping of new CNN architectures. At the same time, it supports efficient computation on CPU and GPU, allowing to train complex models on large datasets such as ImageNet ILSVRC containing millions of training examples

Patterns and textures are key characteristics of many natural objects: a shirt can be striped, the wings of a butterfly can be veined, and the skin of an animal can be scaly. Aiming at supporting this dimension in image understanding, we address the problem of describing textures with semantic attributes. We identify a vocabulary of forty-seven texture terms and use them to describe a large dataset of patterns collected "in the wild". The resulting Describable Textures Dataset (DTD) is a basis to seek the best representation for recognizing describable texture attributes in images. We port from object recognition to texture recognition the Improved Fisher Vector (IFV) and Deep Convolutional-network Activation Features (DeCAF), and show that surprisingly, they both outperform specialized texture descriptors not only on our problem, but also in established material recognition datasets. We also show that our describable attributes are excellent texture descriptors, transferring between datasets and tasks, in particular, combined with IFV and DeCAF, they significantly outperform the state-of-the-art by more than 10% on both FMD and KTH-TIPS-2b benchmarks. We also demonstrate that they produce intuitive descriptions of materials and Internet images.

Our goal is to learn a compact, discriminative vector representation of a face track, suitable for the face recognition tasks of verification and classification. To this end, we propose a novel face track descriptor, based on the Fisher Vector representation, and demonstrate that it has a number of favourable properties. First, the descriptor is suitable for tracks of both frontal and profile faces, and is insensitive to their pose. Second, the descriptor is compact due to discriminative dimensionality reduction, and it can be further compressed using binarization. Third, the descriptor can be computed quickly (using hard quantization) and its compact size and fast computation render it very suitable for large scale visual repositories. Finally, the descriptor demonstrates good generalization when trained on one dataset and tested on another, reflecting its tolerance to the dataset bias. In the experiments we show that the descriptor exceeds the state of the art on both face verification task (YouTube Faces without outside training data, and INRIA-Buffy benchmarks), and face classification task (using the Oxford-Buffy dataset).

We study the problem of understanding objects in detail, intended as recognizing a wide array of fine-grained object attributes. To this end, we introduce a dataset of 7, 413 airplanes annotated in detail with parts and their attributes, leveraging images donated by airplane spotters and crowd-sourcing both the design and collection of the detailed annotations. We provide a number of insights that should help researchers interested in designing fine-grained datasets for other basic level categories. We show that the collected data can be used to study the relation between part detection and attribute prediction by diagnosing the performance of classifiers that pool information from different parts of an object. We note that the prediction of certain attributes can benefit substantially from accurate part detection. We also show that, differently from previous results in object detection, employing a large number of part templates can improve detection accuracy at the expenses of detection speed. We finally propose a coarse-to-fine approach to speed up detection through a hierarchical cascade algorithm.

The objective of this work is to learn descriptors suitable for the sparse feature detectors used in viewpoint invariant matching. We make a number of novel contributions towards this goal. First, it is shown that learning the pooling regions for the descriptor can be formulated as a convex optimisation problem selecting the regions using sparsity. Second, it is shown that descriptor dimensionality reduction can also be formulated as a convex optimisation problem, using Mahalanobis matrix nuclear norm regularisation. Both formulations are based on discriminative large margin learning constraints. As the third contribution, we evaluate the performance of the compressed descriptors, obtained from the learnt real-valued descriptors by binarisation. Finally, we propose an extension of our learning formulations to a weakly supervised case, which allows us to learn the descriptors from unannotated image collections. It is demonstrated that the new learning methods improve over the state of the art in descriptor learning on the annotated local patches data set of Brown et al. and unannotated photo collections of Philbin et al.

Several recent papers on automatic face verification have significantly raised the performance bar by developing novel, specialised representations that outperform standard features such as SIFT for this problem. This paper makes two contributions: first, and somewhat surprisingly, we show that Fisher vectors on densely sampled SIFT features, i.e. an off-the-shelf object recognition representation, are capable of achieving state-of-the-art face verification performance on the challenging “Labeled Faces in the Wild” benchmark; second, since Fisher vectors are very high dimensional, we show that a compact descriptor can be learnt from them using discriminative metric learning. This compact descriptor has a better recognition accuracy and is very well suited to large scale identification tasks.

The automatic discovery of distinctive parts for an object or scene class is challenging since it requires simultaneously to learn the part appearance and also to identify the part occurrences in images. In this paper, we propose a simple, efficient, and effective method to do so. We address this problem by learning parts incrementally, starting from a single part occurrence with an Exemplar SVM. In this manner, additional part instances are discovered and aligned reliably before being considered as training examples. We also propose entropy-rank curves as a means of evaluating the distinctiveness of parts shareable between categories and use them to select useful parts out of a set of candidates. We apply the new representation to the task of scene categorisation on the MIT Scene 67 benchmark. We show that our method can learn parts which are significantly more informative and for a fraction of the cost, compared to previous part-learning methods such as Singh et al. [28]. We also show that a well constructed bag of words or Fisher vector model can substantially outperform the previous state-of-the-art classification performance on this data.

As massively parallel computations have become broadly available with modern GPUs, deep architectures trained on very large datasets have risen in popularity. Discriminatively trained convolutional neural networks, in particular, were recently shown to yield state-of-the-art performance in challenging image classification benchmarks such as ImageNet. However, elements of these architectures are similar to standard hand-crafted representations used in computer vision. In this paper, we explore the extent of this analogy, proposing a version of the state-of-the-art Fisher vector image encoding that can be stacked in multiple layers. This architecture significantly improves on standard Fisher vectors, and obtains competitive results with deep convolutional networks at a significantly smaller computational cost. Our hybrid architecture allows us to measure the performance improvement brought by a deeper image classification pipeline, while staying in the realms of conventional SIFT features and FV encodings.

We investigate the fine grained object categorization problem of determining the breed of animal from an image. To this end we introduce a new annotated dataset of pets covering 37 different breeds of cats and dogs. The visual problem is very challenging as these animals, particularly cats, are very deformable and there can be quite subtle differences between the breeds. We make a number of contributions: first, we introduce a model to classify a pet breed automatically from an image. The model combines shape, captured by a deformable part model detecting the pet face, and appearance, captured by a bag-of-words model that describes the pet fur. Fitting the model involves automatically segmenting the animal in the image. Second, we compare two classification approaches: a hierarchical one, in which a pet is first assigned to the cat or dog family and then to a breed, and a flat one, in which the breed is obtained directly. We also investigate a number of animal and image orientated spatial layouts. These models are very good: they beat all previously published results on the challenging ASIRRA test (cat vs dog discrimination). When applied to the task of discriminating the 37 different breeds of pets, the models obtain an average accuracy of about 59%, a very encouraging result considering the difficulty of the problem.

Efficient learning with non-linear kernels is often based on extracting features from the data that “linearise” the kernel. While most constructions aim at obtaining low-dimensional and dense features, in this work we explore high-dimensional and sparse ones. We give a method to compute sparse features for arbitrary kernels, re-deriving as a special case a popular map for the intersection kernel and extending it to arbitrary additive kernels. We show that bundle optimisation methods can handle efficiently these sparse features in learning. As an application, we show that product quantisation can be interpreted as a sparse feature encoding, and use this to significantly accelerate learning with this technique. We demonstrate these ideas on image classification with Fisher kernels and object detection with deformable part models on the challenging PASCAL VOC data, obtaining five to ten-fold speed-ups as well as reducing memory use by an order of magnitude.

Large scale nonlinear support vector machines (SVMs) can be approximated by linear ones using a suitable feature map. The linear SVMs are in general much faster to learn and evaluate (test) than the original nonlinear SVMs. This work introduces explicit feature maps for the additive class of kernels, such as the intersection, Hellinger's, and χ2 kernels, commonly used in computer vision, and enables their use in large scale problems. In particular, we: 1) provide explicit feature maps for all additive homogeneous kernels along with closed form expression for all common kernels; 2) derive corresponding approximate finite-dimensional feature maps based on a spectral analysis; and 3) quantify the error of the approximation, showing that the error is independent of the data dimension and decays exponentially fast with the approximation order for selected kernels such as χ2. We demonstrate that the approximations have indistinguishable performance from the full kernels yet greatly reduce the train/test times of SVMs. We also compare with two other approximation methods: Nystrom's approximation of Perronnin et al., which is data dependent, and the explicit map of Maji and Berg for the intersection kernel, which, as in the case of our approximations, is data independent. The approximations are evaluated on a number of standard data sets, including Caltech-101, Daimler-Chrysler pedestrians, and INRIA pedestrians.

A large number of novel encodings for bag of visual words models have been proposed in the past two years to improve on the standard histogram of quantized local features. Examples include locality-constrained linear encoding [23], improved Fisher encoding [17], super vector encoding [27], and kernel codebook encoding [20]. While several authors have reported very good results on the challenging PASCAL VOC classification data by means of these new techniques, differences in the feature computation and learning algorithms, missing details in the description of the methods, and different tuning of the various components, make it impossible to compare directly these methods and hard to reproduce the results reported. This paper addresses these shortcomings by carrying out a rigorous evaluation of these new techniques by: (1) fixing the other elements of the pipeline (features, learning, tuning); (2) disclosing all the implementation details, and (3) identifying both those aspects of each method which are particularly important to achieve good performance, and those aspects which are less critical. This allows a consistent comparative analysis of these encoding methods. Several conclusions drawn from our analysis cannot be inferred from the original publications.

We present a method that can dramatically accelerate object detection with part based models. The method is based on the observation that the cost of detection is likely to be dominated by the cost of matching each part to the image, and not by the cost of computing the optimal configuration of the parts as commonly assumed. Therefore accelerating detection requires minimizing the number of part-to-image comparisons. To this end we propose a multiple-resolutions hierarchical part based model and a corresponding coarse-to-fine inference procedure that recursively eliminates from the search space unpromising part placements. The method yields a ten-fold speedup over the standard dynamic programming approach and is complementary to the cascade-of-parts approach of. Compared to the latter, our method does not have parameters to be determined empirically, which simplifies its use during the training of the model. Most importantly, the two techniques can be combined to obtain a very significant speedup, of two orders of magnitude in some cases. We evaluate our method extensively on the PASCAL VOC and INRIA datasets, demonstrating a very high increase in the detection speed with little degradation of the accuracy.

Template-based object detectors such as the deformable parts model of Felzenszwalb et al. [11] achieve state-of-the-art performance for a variety of object categories, but are still outperformed by simpler bag-of-words models for highly flexible objects such as cats and dogs. In these cases we propose to use the template-based model to detect a distinctive part for the class, followed by detecting the rest of the object via segmentation on image specific information learnt from that part. This approach is motivated by two observations: (i) many object classes contain distinctive parts that can be detected very reliably by template-based detectors, whilst the entire object cannot; (ii) many classes (e.g. animals) have fairly homogeneous coloring and texture that can be used to segment the object once a sample is provided in an image. We show quantitatively that our method substantially outperforms whole-body template-based detectors for these highly deformable object categories, and indeed achieves accuracy comparable to the state-of-the-art on the PASCAL VOC competition, which includes other models such as bag-of-words.

Equivariance and invariance are often desired properties of a computer vision system. However, currently available strategies generally rely on virtual sampling, leaving open the question of how many samples are necessary, on the use of invariant feature representations, which can mistakenly discard information relevant to the vision task, or on the use of latent variable models, which result in non-convex training and expensive inference at test time. We propose here a generalization of structured output SVM regressors that can incorporate equivariance and invariance into a convex training procedure, enabling the incorporation of large families of transformations, while maintaining optimality and tractability. Importantly, test time inference does not require the estimation of latent variables, resulting in highly efficient objective functions. This results in a natural formulation for treating equivariance and invariance that is easily implemented as an adaptation of off-the-shelf optimization software, obviating the need for ad hoc sampling strategies. Theoretical results relating to vicinal risk, and experiments on challenging aerial car and pedestrian detection tasks show the effectiveness of the proposed solution.

Graph cut optimization is one of the standard workhorses of image segmentation since for binary random field representations of the image, it gives globally optimal results and there are efficient polynomial time implementations. Often, the random field is applied over a flat partitioning of the image into non-intersecting elements, such as pixels or super-pixels. In the paper we show that if, instead of a flat partitioning, the image is represented by a hierarchical segmentation tree, then the resulting energy combining unary and boundary terms can still be optimized using graph cut (with all the corresponding benefits of global optimality and efficiency). As a result of such inference, the image gets partitioned into a set of segments that may come from different layers of the tree. We apply this formulation, which we call the pylon model, to the task of semantic segmentation where the goal is to separate an image into areas belonging to different semantic classes. The experiments highlight the advantage of inference on a segmentation tree (over a flat partitioning) and demonstrate that the optimization in the pylon model is able to flexibly choose the level of segmentation across the image. Overall, the proposed system has superior segmentation accuracy on several datasets (Graz-02, Stanford background) compared to previously suggested approaches.

Kernel methods yield state-of-the-art performance in certain applications such as image classification and object detection. However, large scale problems require machine learning techniques of at most linear complexity and these are usually limited to linear kernels. This unfortunately rules out gold-standard kernels such as the generalized RBF kernels (e.g. exponential-c 2 ). Recently, Maji and Berg [13] and Vedaldi and Zisserman [20] proposed explicit feature maps to approximate the additive kernels (intersection, c 2 , etc.) by linear ones, thus enabling the use of fast machine learning technique in a non-linear context. An analogous technique was proposed by Rahimi and Recht [14] for the translation invariant RBF kernels. In this paper, we complete the construction and combine the two techniques to obtain explicit feature maps for the generalized RBF kernels. Furthermore, we investigate a learning method using l 1 regularization to encourage sparsity in the final vector representation, and thus reduce its dimension. We evaluate this technique on the VOC 2007 detection challenge, showing when it can improve on fast additive kernels, and the trade-offs in complexity and accuracy.

Maji and Berg have recently introduced an explicit feature map approximating the intersection kernel. This enables efficient learning methods for linear kernels to be applied to the non-linear intersection kernel, expanding the applicability of this model to much larger problems. In this paper we generalize this idea, and analyse a large family of additive kernels, called homogeneous, in a unified framework. The family includes the intersection, Hellinger's, and χ2kernels commonly employed in computer vision. Using the framework we are able to: (i) provide explicit feature maps for all homogeneous additive kernels along with closed form expression for all common kernels; (ii) derive corresponding approximate finite-dimensional feature maps based on the Fourier sampling theorem; and (iii) quantify the extent of the approximation. We demonstrate that the approximations have indistinguishable performance from the full kernel on a number of standard datasets, yet greatly reduce the train/test times of SVM implementations. We show that the χ2kernel, which has been found to yield the best performance in most applications, also has the most compact feature representation. Given these train/test advantages we are able to obtain a significant performance improvement over current state of the art results based on the intersection kernel.

VLFeat is an open and portable library of computer vision algorithms. It aims at facilitating fast prototyping and reproducible research for computer vision scientists and students. It includes rigorous implementations of common building blocks such as feature detectors, feature extractors, (hierarchical) k-means clustering, randomized kd-tree matching, and super-pixelization. The source code and interfaces are fully documented. The library integrates directly with MATLAB, a popular language for computer vision research.

A standard approach to learning object category detectors is to provide strong supervision in the form of a region of interest (ROI) specifying each instance of the object in the training images. In this work are goal is to learn from heterogeneous labels, in which some images are only weakly supervised, specifying only the presence or absence of the object or a weak indication of object location, whilst others are fully annotated. To this end we develop a discriminative learning approach and make two contributions: (i) we propose a structured output formulation for weakly annotated images where full annotations are treated as latent variables; and (ii) we propose to optimize a ranking objective function, allowing our method to more effectively use negatively labeled images to improve detection average precision performance. The method is demonstrated on the benchmark INRIA pedestrian detection dataset of Dalal and Triggs and the PASCAL VOC dataset, and it is shown that for a significant proportion of weakly supervised images the performance achieved is very similar to the fully supervised (state of the art) results.

We propose a method to identify and localize object classes in images. Instead of operating at the pixel level, we advocate the use of superpixels as the basic unit of a class segmentation or pixel localization scheme. To this end, we construct a classifier on the histogram of local features found in each superpixel. We regularize this classifier by aggregating histograms in the neighborhood of each superpixel and then refine our results further by using the classifier in a conditional random field operating on the superpixel graph. Our proposed method exceeds the previously published state-of-the-art on two challenging datasets: Graz-02 and the PASCAL VOC 2007 Segmentation Challenge.

Our objective is to obtain a state-of-the art object category detector by employing a state-of-the-art image classifier to search for the object in all possible image sub-windows. We use multiple kernel learning of Varma and Ray (ICCV 2007) to learn an optimal combination of exponential χ2kernels, each of which captures a different feature channel. Our features include the distribution of edges, dense and sparse visual words, and feature descriptors at different levels of spatial organization. Such a powerful classifier cannot be tested on all image sub-windows in a reasonable amount of time. Thus we propose a novel three-stage classifier, which combines linear, quasi-linear, and non-linear kernel SVMs. We show that increasing the non-linearity of the kernels increases their discriminative power, at the cost of an increased computational complexity. Our contributions include (i) showing that a linear classifier can be evaluated with a complexity proportional to the number of sub-windows (independent of the sub-window area and descriptor dimension); (ii) a comparison of three efficient methods of proposing candidate regions (including the jumping window classifier of Chum and Zisserman (CVPR 2007) based on proposing windows from scale invariant features); and (Hi) introducing overlap-recall curves as a mean to compare and optimize the performance of the intermediate pipeline stages. The method is evaluated on the PASCAL Visual Object Detection Challenge, and exceeds the performances of previously published methods for most of the classes.

We develop a structured output model for object category detection that explicitly accounts for alignment, multiple aspects and partial truncation in both training and inference. The model is formulated as large margin learning with latent variables and slack rescaling, and both training and inference are computationally efficient. We make the following contributions: (i) we note that extending the Structured Output Regression formulation of Blaschko and Lampert (ECCV 2008) to include a bias term significantly improves performance; (ii) that alignment (to account for small rotations and anisotropic scalings) can be included as a latent variable and efficiently determined and implemented; (iii) that the latent variable extends to multiple aspects (e.g. left facing, right facing, front) with the same formulation; and (iv), most significantly for performance, that truncated and truncated instances can be included in both training and inference with an explicit truncation mask. We demonstrate the method by training and testing on the PASCAL VOC 2007 data set -- training includes the truncated examples, and in testing object instances are detected at multiple scales, alignments, and with significant truncations.

Joint data alignment is often regarded as a data simplification process. This idea is powerful and general, but raises two delicate issues. First, one must make sure that the use full information about the data is preserved by the alignment process. This is especially important when data are affected by non-invertible transformations, such as those originating from continuous domain deformations in a discrete image lattice. We propose a formulation that explicitly avoids this pitfall. Second, one must choose an appropriate measure of data complexity. We show that standard concepts such as entropy might not be optimal for the task, and we propose alternative measures that reflect the regularity of the codebook space. We also propose a novel and efficient algorithm that allows joint alignment of a large number of samples (tens of thousands of image patches), and does not rely on the assumption that pixels are independent. This is done for the case where the data is postulated to live in an affine subspaces of the embedding space of the raw data. We apply our scheme to learn sparse bases for natural images that discount domain deformations and hence significantly decrease the complexity of codebooks while maintaining the same generative power.

The popular bag-of-features representation for object recognition collects signatures of local image patches and discards spatial information. Some have recently attempted to at least partially overcome this limitation, for instance by ldquospatial pyramidsrdquo and ldquoproximityrdquo kernels. We introduce the general formalism of ldquorelaxed matching kernelsrdquo (RMKs) that includes such approaches as special cases, allow us to derive useful general properties of these kernels, and to introduce new ones. As an example, we introduce a kernel based on matching graphs of features and one based on matching information-compressed features. We show that all RMKs are competitive and outperform in several cases recently published state-of-the-art results on standard datasets. However, we also show that a proper implementation of a baseline bag-of-features algorithm can be extremely competitive, and outperform the other methods in some cases.

It is well-known that forward motion induces a large number of local minima in the instantaneous least-squares reprojection error. This is caused in part by singularities in the error landscape around the forward direction, and presents a challenge in using existing algorithms for structure-from-motion in autonomous navigation applications. In this paper we prove that imposing a bound on the reconstructed depth of the scene makes the least-squares re-projection error continuous. This has implications for autonomous navigation, as it suggests simple modifications for existing algorithms to minimize the effects of local minima in forward translation.

In the task of visual object categorization, semantic context can play the very important role of reducing ambiguity in objects' visual appearance. In this work we propose to incorporate semantic object context as a post-processing step into any off-the-shelf object categorization model. Using a conditional random field (CRF) framework, our approach maximizes object label agreement according to contextual relevance. We compare two sources of context: one learned from training data and another queried from Google Sets. The overall performance of the proposed framework is evaluated on the PASCAL and MSRC datasets. Our findings conclude that incorporating context into object categorization greatly improves categorization accuracy.

In this paper we present a novel boosting algorithm for supervised learning that incorporates invariance to data transformations and has high generalization capabilities. While one can incorporate invariance by adding virtual samples to the data (e.g., by jittering), we adopt a much more efficient strategy and work along the lines of vicinal risk minimization and tangent distance methods. As in vicinal risk minimization, we incorporate invariance to data by applying anisotropic smoothing along the directions of invariance. Moreover, as in tangent distance methods, we provide a simple local approximation to such directions, thus obtaining an efficient computational scheme. We also show that it is possible to automatically design optimal weak classifiers by using gradient descent. To increase efficiency at run time, such optimal weak classifiers are projected on a Haar basis. This results in designing strong classifiers that are more computationally efficient than in the case of exhaustive search. For illustration and validation purposes, we demonstrate the novel algorithm both on synthetic and on real data sets that are publicly available.

We present a technique to adapt the domain of local features through the matching process to augment their discriminative power. We start with local affine features selected and normalized independently in training and test images, and jointly expand their domain as part of the correspondence process, akin to a (non-rigid) registration task that yields a (multi-view) segmentation of the object of interest from clutter, including the detection of occlusions. We show how our growth process can be used to validate putative affine matches, to match a given "template" (an image of an object without clutter) to a cluttered and partially occluded image, and to match two images that contain the same unknown object in different clutter under different occlusions (unsupervised object detection).

Image Congealing (IC) is a non-parametric method for the joint alignment of a col- lection of images affected by systematic and unwanted deformations. The method attempts to undo the deformations by minimizing a measure of complexity of the image ensemble, such as the averaged per-pixel entropy. This enables alignment without an explicit model of the aligned dataset as required by other methods (e.g. transformed component analysis). While IC is simple and general, it may intro- duce degenerate solutions when the transformations allow minimizing the com- plexity of the data by collapsing them to a constant. Such solutions need to be explicitly removed by regularization. In this paper we propose an alternative formulation which solves this regulariza- tion issue on a more principled ground. We make the simple observation that alignment should simplify the data while preserving the useful information car- ried by them. Therefore we trade off fidelity and complexity of the aligned en- semble rather than minimizing the complexity alone. This eliminates the need for an explicit regularization of the transformations, and has a number of other useful properties such as noise suppression. We show the modeling and computa- tional benefits of the approach to the some of the problems on which IC has been demonstrated.

We propose an algorithm to perform causal inference of the state of a dynamical model when the measurements are corrupted by outliers. While the optimal (maximum-likelihood) solution has doubly exponential complexity due to the combinatorial explosion of possible choices of inliers, we exploit the structure of the problem to design a sampling-based algorithm that has constant complexity. We derive our algorithm from the equations of the optimal filter, which makes our approximation explicit. Our work is motivated by real-time tracking and the estimation of structure from motion (SFM). We test our algorithm for on-line outlier rejection both for tracking and for SFM. We show that our approach can tolerate a large proportion of outliers, whereas previous causal robust statistical inference methods failed with less than half as many. Our work can be thought of as the extension of random sample consensus algorithms to dynamic data, or as the implementation of pseudo-Bayesian filtering algorithms in a sampling framework.

Most current local feature detectors/descriptors implicitly assume that the scene is (locally) planar, an assumption that is violated at surface discontinuities. We show that this restriction is, at least in theory, unnecessary, as one can construct local features that are viewpoint-invariant for generic non-planar scenes. However, we show that any such feature necessarily sacrifices shape information, in the sense of being non shape-discriminative. Finally, we show that if viewpoint is factored out as part of the matching process, rather than explicitly in the representation, then shape is discriminative indeed. We illustrate our theoretical results empirically by showing that, even for simple scenes, current affine descriptors fail where even a naive 3-D viewpoint invariant succeeds in matching