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/.

Continual learning (CL) promises to allow neural networks to learn from continuous streams of inputs, instead of IID (independent and identically distributed) sampling, which requires random access to a full dataset. This would allow for much smaller storage requirements and self-sufficiency of deployed systems that cope with natural distribution shifts, similarly to biological learning. We focus on video CL employing a rehearsal-based approach, which reinforces past samples from a memory buffer. We posit that part of the reason why practical video CL is challenging is the high memory requirements of video, further exacerbated by long-videos and continual streams, which are at odds with the common rehearsal-buffer size constraints. To address this, we propose to use compressed vision, i.e. store video codes (embeddings) instead of raw inputs, and train a video classifier by IID sampling from this rolling buffer. Training a video compressor online (so not depending on any pre-trained networks) means that it is also subject to catastrophic forgetting. We propose a scheme to deal with this forgetting by refreshing video codes, which requires careful decompression with a previous version of the network and recompression with a new one. We name our method Continually Refreshed Amodal Memory (CRAM). We expand current video CL benchmarks to large-scale settings, namely EpicKitchens-100 and Kinetics-700, storing thousands of relatively long videos in under 2 GB, and demonstrate empirically that our video CL method outperforms prior art with a significantly reduced memory footprint.

We tackle the problem of localizing traffic cameras within a 3D reference map and propose a novel image-to-point cloud registration (I2P) method, TrafficLoc, in a coarse-tofine matching fashion. To overcome the lack of large-scale real-world intersection datasets, we first introduce Carla Intersection, a new simulated dataset with 75 urban and rural intersections in Carla. We find that current I2P methods struggle with cross-modal matching under large viewpoint differences, especially at traffic intersections. TrafficLoc thus employs a novel Geometry-guided Attention Loss (GAL) to focus only on the corresponding geometric regions under different viewpoints during 2D-3D feature fusion. To address feature inconsistency in paired image patch-point groups, we further propose Inter-intra Contrastive Learning (ICL) to enhance separating 2D patch/3D group features within each intra-modality and introduce Dense Training Alignment (DTA) with soft-argmax for improving position regression. Extensive experiments show our TrafficLoc greatly improves the performance over the SOTA I2P methods (up to 86%) on Carla Intersection and generalizes well to real-world data. TrafficLoc also achieves new SOTA performance on KITTI and NuScenes datasets, demonstrating the superiority across both in-vehicle and traffic cameras. Our project page is publicly available at https://tum-luk.github.io/projects/trafficloc/.

Broadly intelligent agents should form task-specific abstractions that selectively expose the essential elements of a task, while abstracting away the complexity of the raw sensorimotor space. In this work, we present Neuro-Symbolic Predicates, a first-order abstraction language that combines the strengths of symbolic and neural knowledge representations. We outline an online algorithm for inventing such predicates and learning abstract world models. We compare our approach to hierarchical reinforcement learning, vision-language model planning, and symbolic predicate invention approaches, on both in- and out-of-distribution tasks across five simulated robotic domains. Results show that our approach offers better sample complexity, stronger out-of-distribution generalization, and improved interpretability.

Past work has studied the effects of fine-tuning on large language models' (LLMs) overall performance on certain tasks. However, a way to quantitatively and systematically analyze its effect on individual outputs is still lacking. In this work, we propose a new method for measuring the contribution that fine-tuning makes to individual LLM responses, assuming access to the original pre-trained model. Our method takes into account the model's intermediate hidden states, giving a more fine-grained insight into the effects of fine-tuning than a simple comparison of the final outputs of pre-trained and fine-tuned models. We introduce and theoretically analyze an exact decomposition of any fine-tuned LLM into a pre-training component and a fine-tuning component. Empirically, we find that one can steer model behavior and performance by up- or down-scaling the fine-tuning component during the forward pass. Motivated by this finding and our theoretical analysis, we define the Tuning Contribution ($\mathrm{TuCo}$) in terms of the ratio of the magnitudes fine-tuning component and the pre-training component. We find that three prominent adversarial attacks on LLMs circumvent safety measures in a way that reduces the Tuning Contribution, and that $\mathrm{TuCo}$ is consistently lower on prompts where the attacks succeed compared to ones where they do not. This suggests that attenuating the effect of fine-tuning on model outputs plays a role in the success of these attacks. In short, $\mathrm{TuCo}$ enables the quantitative study of how fine-tuning influences model behavior and safety, and vice-versa.

Extracting point correspondences from two or more views of a scene is a fundamental computer vision problem with particular importance for relative camera pose estimation and structure-from-motion. Existing local feature matching approaches, trained with correspondence supervision on large-scale datasets, obtain highly-accurate matches on the test sets. However, they do not generalise well to new datasets with different characteristics to those they were trained on, unlike classic feature extractors. Instead, they require finetuning, which assumes that ground-truth correspondences or ground-truth camera poses and 3D structure are available. We relax this assumption by removing the requirement of 3D structure, e.g., depth maps or point clouds, and only require camera pose information, which can be obtained from odometry. We do so by replacing correspondence losses with epipolar losses, which encourage putative matches to lie on the associated epipolar line. While weaker than correspondence supervision, we observe that this cue is sufficient for finetuning existing models on new data. We then further relax the assumption of known camera poses by using pose estimates in a novel bootstrapping approach. We evaluate on highly challenging datasets, including an indoor drone dataset and an outdoor smartphone camera dataset, and obtain state-of-the-art results without strong supervision.

We tackle the problem of 3D point cloud localization based on a few natural linguistic descriptions and introduce a novel neural network, Text2Loc, that fully interprets the semantic relationship between points and text. Text2Loc follows a coarse-to-fine localization pipeline: text-submap global place recognition, followed by fine localization. In global place recognition, relational dynamics among each textual hint are captured in a hierarchical transformer with max-pooling (HTM), whereas a balance between positive and negative pairs is maintained using text-submap contrastive learning. Moreover, we propose a novel matching-free fine localization method to further refine the location predictions, which completely removes the need for complicated text-instance matching and is lighter, faster, and more accurate than previous methods. Extensive experiments show that Text2Loc improves the localization accuracy by up to 2 × over the state-of-the-art on the KITTI360Pose dataset. Our project page is publicly available at https://yan-xia.github.io/projects/text2loc/.

Developing generalizable manipulation skills is a core challenge in embodied AI. This includes generalization across diverse task configurations, encompassing variations in object shape, density, friction coefficient, and external disturbances such as forces applied to the robot. Rapid Motor Adaptation (RMA) offers a promising solution to this challenge. It posits that essential hidden variables influencing an agent's task performance, such as object mass and shape, can be effectively inferred from the agent's action and proprioceptive history. Drawing inspiration from RMA in locomotion and in-hand rotation, we use depth perception to develop agents tailored for rapid motor adaptation in a variety of manipulation tasks. We evaluated our agents on four challenging tasks from the Maniskill2 benchmark, namely pick-and-place operations with hundreds of objects from the YCB and EGAD datasets, peg insertion with precise position and orientation, and operating a variety of faucets and handles, with customized environment variations. Empirical results demonstrate that our agents surpass state-of-the-art methods like automatic domain randomization and vision-based policies, obtaining better generalization performance and sample efficiency.

Video databases from the internet are a valuable source of text-audio retrieval datasets. However, given that sound and vision streams represent different "views" of the data, treating visual descriptions as audio descriptions is far from optimal. Even if audio class labels are present, they commonly are not very detailed, making them unsuited for text-audio retrieval. To exploit relevant audio information from video-text datasets, we introduce a methodology for generating audio-centric descriptions using Large Language Models (LLMs). In this work, we consider the egocentric video setting and propose three new text-audio retrieval benchmarks based on the EpicMIR and EgoMCQ tasks, and on the EpicSounds dataset. Our approach for obtaining audio-centric descriptions gives significantly higher zero-shot performance than using the original visual-centric descriptions. Furthermore, we show that using the same prompts, we can successfully employ LLMs to improve the retrieval on EpicSounds, compared to using the original audio class labels of the dataset. Finally, we confirm that LLMs can be used to determine the difficulty of identifying the action associated with a sound.

AI agents are commonly trained with large datasets of demonstrations of human behavior. However, not all behaviors are equally safe or desirable. Desired characteristics for an AI agent can be expressed by assigning desirability scores, which we assume are not assigned to individual behaviors but to collective trajectories. For example, in a dataset of vehicle interactions, these scores might relate to the number of incidents that occurred. We first assess the effect of each individual agent's behavior on the collective desirability score, e.g., assessing how likely an agent is to cause incidents. This allows us to selectively imitate agents with a positive effect, e.g., only imitating agents that are unlikely to cause incidents. To enable this, we propose the concept of an agent's \textit{Exchange Value}, which quantifies an individual agent's contribution to the collective desirability score. The Exchange Value is the expected change in desirability score when substituting the agent for a randomly selected agent. We propose additional methods for estimating Exchange Values from real-world datasets, enabling us to learn desired imitation policies that outperform relevant baselines. The project website can be found at https://tinyurl.com/select-to-perfect.

Autonomous agents deployed in the real world need to be robust against adversarial attacks on sensory inputs. Robustifying agent policies requires anticipating the strongest attacks possible. We demonstrate that existing observation-space attacks on reinforcement learning agents have a common weakness: while effective, their lack of information-theoretic detectability constraints makes them \textit{detectable} using automated means or human inspection. Detectability is undesirable to adversaries as it may trigger security escalations. We introduce \textit{\eattacks{}}, a novel form of adversarial attack on sequential decision-makers that is both effective and of $\epsilon-$bounded statistical detectability. We propose a novel dual ascent algorithm to learn such attacks end-to-end. Compared to existing attacks, we empirically find \eattacks{} to be significantly harder to detect with automated methods, and a small study with human participants\footnote{IRB approval under reference R84123/RE001} suggests they are similarly harder to detect for humans. Our findings suggest the need for better anomaly detectors, as well as effective hardware- and system-level defenses. The project website can be found at https://tinyurl.com/illusory-attacks.

Recent advancements in machine learning have fueled research on multimodal tasks, such as for instance text-to-video and text-to-audio retrieval. These tasks require models to understand the semantic content of video and audio data, including objects, and characters. The models also need to learn spatial arrangements and temporal relationships. In this work, we analyse the temporal ordering of sounds, which is an understudied problem in the context of text-to-audio retrieval. In particular, we dissect the temporal understanding capabilities of a state-of-the-art model for text-to-audio retrieval on the AudioCaps and Clotho datasets. Additionally, we introduce a synthetic text-audio dataset that provides a controlled setting for evaluating temporal capabilities of recent models. Lastly, we present a loss function that encourages text-audio models to focus on the temporal ordering of events. Code and data are available at https://github.com/oncescuandreea/DTU_text_audio.

Image feature extractors are rendered substantially more useful if different views of the same 3D location yield similar features while still being distinct from other locations. A feature extractor that achieves this goal even under significant viewpoint changes must recognise not just semantic categories in a scene, but also understand how different objects relate to each other in three dimensions. Existing work addresses this task by posing it as a patch retrieval problem, training the extracted features to facilitate retrieval of all image patches that project from the same 3D location. However, this approach uses a loss formulation that requires substantial memory and computation resources, limiting its applicability for large-scale training. We present a method for memory-efficient learning of location-consistent features that reformulates and approximates the smooth average precision objective. This novel loss function enables improvements in memory efficiency by three orders of magnitude, mitigating a key bottleneck of previous methods and allowing much larger models to be trained with the same computational resources. We showcase the improved location consistency of our trained feature extractor directly on a multi-view consistency task, as well as the downstream task of scene-stable panoptic segmentation, significantly outperforming previous state-of-the-art.

Unsupervised object detection using deep neural networks is typically a difficult problem with few to no guarantees about the learned representation. In this work we present the first unsupervised object detection method that is theoretically guaranteed to recover the true object positions up to quantifiable small shifts. We develop an unsupervised object detection architecture and prove that the learned variables correspond to the true object positions up to small shifts related to the encoder and decoder receptive field sizes, the object sizes, and the widths of the Gaussians used in the rendering process. We perform detailed analysis of how the error depends on each of these variables and perform synthetic experiments validating our theoretical predictions up to a precision of individual pixels. We also perform experiments on CLEVR-based data and show that, unlike current SOTA object detection methods (SAM, CutLER), our method's prediction errors always lie within our theoretical bounds. We hope that this work helps open up an avenue of research into object detection methods with theoretical guarantees.

Transformers are powerful visual learners, in large part due to their conspicuous lack of manually-specified priors. This flexibility can be problematic in tasks that involve multiple-view geometry, due to the near-infinite possible variations in 3D shapes and viewpoints (requiring flexibility), and the precise nature of projective geometry (obeying rigid laws). To resolve this conundrum, we propose a “light touch” approach, guiding visual Transformers to learn multiple-view geometry but allowing them to break free when needed. We achieve this by using epipolar lines to guide the Transformer's cross-attention maps during training, penalizing attention values outside the epipolar lines and encouraging higher attention along these lines since they contain geometrically plausible matches. Unlike previous methods, our proposal does not require any camera pose information at test-time. We focus on pose-invariant object instance retrieval, where standard Transformer networks struggle, due to the large differences in viewpoint between query and retrieved images. Experimentally, our method outperforms state-of-the-art approaches at object retrieval, without needing pose information at test-time.

Place recognition based on point clouds (LiDAR) is an important component for autonomous robots or self-driving vehicles. Current SOTA performance is achieved on accumulated LiDAR submaps using either point-based or voxel-based structures. While voxel-based approaches nicely integrate spatial context across multiple scales, they do not exhibit the local precision of point-based methods. As a result, existing methods struggle with fine-grained matching of subtle geometric features in sparse single-shot Li-DAR scans. To overcome these limitations, we propose CASSPR as a method to fuse point-based and voxel-based approaches using cross attention transformers. CASSPR leverages a sparse voxel branch for extracting and aggregating information at lower resolution and a point-wise branch for obtaining fine-grained local information. CASSPR uses queries from one branch to try to match structures in the other branch, ensuring that both extract self-contained descriptors of the point cloud (rather than one branch dominating), but using both to inform the out-put global descriptor of the point cloud. Extensive experiments show that CASSPR surpasses the state-of-the-art by a large margin on several datasets (Oxford RobotCar, TUM, USyd). For instance, it achieves AR@1 of 85.6% on the TUM dataset, surpassing the strongest prior model by ~15%. Our code is publicly available.1

An important challenge for autonomous agents such as robots is to maintain a spatially and temporally consistent model of the world. It must be maintained through occlusions, previously-unseen views, and long time horizons (e.g., loop closure and re-identification). It is still an open question how to train such a versatile neural representation without supervision. We start from the idea that the training objective can be framed as a patch retrieval problem: given an image patch in one view of a scene, we would like to retrieve (with high precision and recall) all patches in other views that map to the same real-world location. One drawback is that this objective does not promote reusability of features: by being unique to a scene (achieving perfect precision/recall), a representation will not be useful in the context of other scenes. We find that it is possible to balance retrieval and reusability by constructing the retrieval set carefully, leaving out patches that map to far-away locations. Similarly, we can easily regulate the scale of the learned features (e.g., points, objects, or rooms) by adjusting the spatial tolerance for considering a retrieval to be positive. We optimize for (smooth) Average Precision (AP), in a single unified ranking-based objective. This objective also doubles as a criterion for choosing landmarks or keypoints, as patches with high AP. We show results creating sparse, multi-scale, semantic spatial maps composed of highly identifiable landmarks, with applications in landmark retrieval, localization, semantic segmentation and instance segmentation.

Standard semantic segmentation models owe their success to curated datasets with a fixed set of semantic categories, without contemplating the possibility of identifying unknown objects from novel categories. Existing methods in outlier detection suffer from a lack of smoothness and objectness in their predictions, due to limitations of the per-pixel classification paradigm. Furthermore, additional training for detecting outliers harms the performance of known classes. In this paper, we explore another paradigm with region-level classification to better segment unknown objects. We show that the object queries in mask classification tend to behave like one vs. all classifiers. Based on this finding, we propose a novel outlier scoring function called RbA by defining the event of being an outlier as being rejected by all known classes. Our extensive experiments show that mask classification improves the performance of the existing outlier detection methods, and the best results are achieved with the proposed RbA. We also propose an objective to optimize RbA using minimal outlier supervision. Further fine-tuning with outliers improves the unknown performance, and unlike previous methods, it does not degrade the inlier performance. Project page: https://kuis-ai.github.io/RbA

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.

Diffusion models have achieved remarkable results in image generation, and have similarly been used to learn high-performing policies in sequential decision-making tasks. Decision-making diffusion models can be trained on lower-quality data, and then be steered with a reward function to generate near-optimal trajectories. We consider the problem of extracting a reward function by comparing a decision-making diffusion model that models low-reward behavior and one that models high-reward behavior; a setting related to inverse reinforcement learning. We first define the notion of a \emph{relative reward function of two diffusion models} and show conditions under which it exists and is unique. We then devise a practical learning algorithm for extracting it by aligning the gradients of a reward function -- parametrized by a neural network -- to the difference in outputs of both diffusion models. Our method finds correct reward functions in navigation environments, and we demonstrate that steering the base model with the learned reward functions results in significantly increased performance in standard locomotion benchmarks. Finally, we demonstrate that our approach generalizes beyond sequential decision-making by learning a reward-like function from two large-scale image generation diffusion models. The extracted reward function successfully assigns lower rewards to harmful images.

The objectives of this work arecross-modal text-audio and audio-text retrieval, in which the goal is to retrieve the audio content from a pool of candidates that best matches a given written description and vice versa. Text-audio retrieval enables users to search large databases through an intuitive interface: they simply issue free-form natural language descriptions of the sound they would like to hear. To study the tasks of text-audio and audio-text retrieval, which have received limited attention in the existing literature, we introduce three challenging new benchmarks. We first construct text-audio and audio-text retrieval benchmarks from theAudioCapsandClothoaudio captioning datasets. Additionally, we introduce theSoundDescsbenchmark, which consists of paired audio and natural language descriptions for a diverse collection of sounds that are complementary to those found inAudioCapsandClotho. We employ these three benchmarks to establish baselines for cross-modal text-audio and audio-text retrieval, where we demonstrate the benefits of pre-training on diverse audio tasks. We hope that our benchmarks will inspire further research into audio retrieval with free-form text queries.

We train embodied neural networks to plan and navigate unseen complex 3D environments, emphasising real-world deployment. Rather than requiring prior knowledge of the agent or environment, the planner learns to model the state transitions and rewards. To avoid the potentially hazardous trial-and-error of reinforcement learning, we focus on differentiable planners such as Value Iteration Networks (VIN), which are trained offline from safe expert demonstrations. Although they work well in small simulations, we address two major limitations that hinder their deployment. First, we observed that current differentiable planners struggle to plan long-term in environments with a high branching complexity. While they should ideally learn to assign low rewards to obstacles to avoid collisions, these penalties are not strong enough to guarantee collision-free operation. We thus impose a structural constraint on the value iteration, which explicitly learns to model impossible actions and noisy motion. Secondly, we extend the model to plan exploration with a limited perspective camera under translation and fine rotations, which is crucial for real robot deployment. Our proposals significantly improve semantic navigation and exploration on several 2D and 3D environments, succeeding in settings that are otherwise challenging for differentiable planners. As far as we know, we are the first to successfully apply them to the difficult Active Vision Dataset, consisting of real images captured from a robot.11Code available: https://github.com/shuishida/calvin

We study how an autonomous agent learns to perform a task from demonstrations in a different domain, such as a different environment or different agent. Such cross-domain imitation learning is required to, for example, train an artificial agent from demonstrations of a human expert. We propose a scalable framework that enables cross-domain imitation learning without access to additional demonstrations or further domain knowledge. We jointly train the learner agent's policy and learn a mapping between the learner and expert domains with adversarial training. We effect this by using a mutual information criterion to find an embedding of the expert's state space that contains task-relevant information and is invariant to domain specifics. This step significantly simplifies estimating the mapping between the learner and expert domains and hence facilitates end-to-end learning. We demonstrate successful transfer of policies between considerably different domains, without extra supervision such as additional demonstrations, and in situations where other methods fail.

We introduce QuerYD, a new large-scale dataset for retrieval and event localisation in video. A unique feature of our dataset is the availability of two audio tracks for each video: the original audio, and a high-quality spoken description of the visual content. The dataset is based on YouDescribe [1], a volunteer project that assists visually-impaired people by attaching voiced narrations to existing YouTube videos. This ever-growing collection of videos contains highly detailed, temporally aligned audio and text annotations. The content descriptions are more relevant than dialogue, and more detailed than previous description attempts, which can be observed to contain many superficial or uninformative descriptions. To demonstrate the utility of the QuerYD dataset, we show that it can be used to train and benchmark strong models for retrieval and event localisation. Data, code and models are made publicly available, and we hope that QuerYD inspires further research on video understanding with written and spoken natural language.

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 consider the task of retrieving audio using free-form natural language queries. To study this problem, which has received limited attention in the existing literature, we introduce challenging new benchmarks for text-based audio retrieval using text annotations sourced from the Audiocaps and Clotho datasets. We then employ these benchmarks to establish baselines for cross-modal audio retrieval, where we demonstrate the benefits of pre-training on diverse audio tasks. We hope that our benchmarks will inspire further research into cross-modal text-based audio retrieval with free-form text queries.

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.

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.

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.

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.

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.

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).

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.

This paper presents a novel Bayesian formulation for aligning faces in unseen images. Our approach revisits the Constrained Local Models (CLM) formulation where an ensemble of local feature detectors are constrained to lie within the subspace spanned by a Point Distribution Model (PDM). Fitting such a model to an image typically involves two main steps: a local search using a detector, obtaining response maps for each landmark (likelihood term) and a global optimization that finds the PDM parameters that jointly maximize all the detections at once. The so-called global optimization can be posed as a Bayesian inference problem, where the posterior distribution of the shape (and pose) parameters can be inferred in a maximum a posteriori (MAP) sense. This work introduces an extended Bayesian global optimization strategy that includes two novel additions: (1) to perform second order updates of the PDM parameters (accounting for their covariance) and (2) to model the underlying dynamics of the shape variations, encoded in the prior term, by using recursive Bayesian estimation. Extensive evaluations were performed against state-of-the-art methods on several standard datasets (IMM, BioID, XM2VTS, LFW and FGNET Talking Face). Results show that the proposed approach significantly increases the fitting performance.

Recently, a particular paradigm [18] in the domain adaptation field has received considerable attention by introducing novel and important insights to the problem. In this case, the source and target domains are represented in the form of subspaces, which are treated as points on the Grassmann manifold. The geodesic curve between them is sampled to obtain intermediate points. Then a classifier is learnt using the projections of the data onto these subspaces. Despite its relevance and popularity, this paradigm [18] contains some limitations. Firstly, in real-world applications, that simple curve (i.e. shortest path) does not provide the necessary flexibility to model the domain shift between the training and testing data sets. Secondly, by using the geodesic curve, we are restricted to only one source domain, which does not allow to take fully advantage of the multiple datasets that are available nowadays. It is then, natural to ask whether this popular concept could be extended to deal with more complex curves and to integrate multi-sources domains. This is a hard problem considering the Riemannian structure of the space, but we propose a mathematically well-founded idea that enables us to solve it. We exploit the geometric insight of rolling maps [30] to compute a spline curve on the Grassmann manifold. The benefits of the proposed idea are demonstrated through several empirical studies on standard datasets. This novel concept allows to explicitly integrate multi-source domains while the previous one [18] uses the mean of all sources. This enables to model better the domain shift and take fully advantage of the training datasets.

The core component of most modern trackers is a discriminative classifier, tasked with distinguishing between the target and the surrounding environment. To cope with natural image changes, this classifier is typically trained with translated and scaled sample patches. Such sets of samples are riddled with redundancies-any overlapping pixels are constrained to be the same. Based on this simple observation, we propose an analytic model for datasets of thousands of translated patches. By showing that the resulting data matrix is circulant, we can diagonalize it with the discrete Fourier transform, reducing both storage and computation by several orders of magnitude. Interestingly, for linear regression our formulation is equivalent to a correlation filter, used by some of the fastest competitive trackers. For kernel regression, however, we derive a new kernelized correlation filter (KCF), that unlike other kernel algorithms has the exact same complexity as its linear counterpart. Building on it, we also propose a fast multi-channel extension of linear correlation filters, via a linear kernel, which we call dual correlation filter (DCF). Both KCF and DCF outperform top-ranking trackers such as Struck or TLD on a 50 videos benchmark, despite running at hundreds of frames-per-second, and being implemented in a few lines of code (Algorithm 1). To encourage further developments, our tracking framework was made open-source.

This paper addresses to the problem of aligning images in unseen faces. The Constrained Local Models (CLM) are popular methods that combine a set of local landmark detectors whose locations are constrained to lie in a subspace spanned by a linear shape model. The CLM fitting is usually based on a two step approach: locally search, using the detectors, producing response maps (likelihood) followed by a global optimization strategy that jointly maximize all detections at once. In this paper, we mainly focus on the first stage: improving the detectors reliability. Usually the local landmarks detectors are far from perfect. Most often are designed to be fast, having a small support region and are learnt from limited data. As consequence, they will suffer from detection ambiguities. Here we propose to improve the detectors performance by considering multiple detection per landmark. In particular, we propose a joint learning of the detectors by clustering of their training data. Afterwards, the multiple likelihoods are combined using a nonlinear fusion approach. The performance evaluation shows that our (extended) approach further increases the fitting performance of the CLM formulation, when compared with recent state-of-the-art methods.

In the past few years there has been a growing interest on geometric frameworks to learn supervised classification models on Riemannian manifolds [32, 28]. A popular framework, valid over any Riemannian manifold, was proposed in [32] for binary classification. Once moving from binary to multi-class classification this paradigm is not valid anymore, due to the spread of multiple positive classes on the manifold [28]. It is then natural to ask whether the multi-class paradigm could be extended to operate on a large class of Riemannian manifolds. We propose a mathematically well-founded classification paradigm that allows to extend the work in [32] to multi-class models, taking into account the structure of the space. The idea is to project all the data from the manifold onto an affine tangent space at a particular point. To mitigate the distortion induced by local diffeomorphisms, we introduce for the first time in the computer vision community a well-founded mathematical concept, so-called Rolling map [22, 17]. The novelty in this alternate school of thought is that the manifold will be firstly rolled (without slipping or twisting) as a rigid body, then the given data is unwrapped onto the affine tangent space, where the classification is performed.

Competitive sliding window detectors require vast training sets. Since a pool of natural images provides a nearly endless supply of negative samples, in the form of patches at different scales and locations, training with all the available data is considered impractical. A staple of current approaches is hard negative mining, a method of selecting relevant samples, which is nevertheless expensive. Given that samples at slightly different locations have overlapping support, there seems to be an enormous amount of duplicated work. It is natural, then, to ask whether these redundancies can be eliminated. In this paper, we show that the Gram matrix describing such data is block-circulant. We derive a transformation based on the Fourier transform that block-diagonalizes the Gram matrix, at once eliminating redundancies and partitioning the learning problem. This decomposition is valid for any dense features and several learning algorithms, and takes full advantage of modern parallel architectures. Surprisingly, it allows training with all the potential samples in sets of thousands of images. By considering the full set, we generate in a single shot the optimal solution, which is usually obtained only after several rounds of hard negative mining. We report speed gains on Caltech Pedestrians and INRIA Pedestrians of over an order of magnitude, allowing training on a desktop computer in a couple of minutes.

This work presents a novel Bayesian formulation for aligning faces in unseen images. Our approach is closely related to Constrained Local Models (CLM) [2] and Active Shape Models (ASM) [6], where an ensemble of local feature detectors are constrained to lie within the subspace spanned by a Point Distribution Model (PDM). Fitting a model to an image typically involves two steps: a local search using a detector, obtaining response maps for each landmark (likelihood term) and a global optimization that finds the PDM parameters that jointly maximize all the detections. The global optimization can be seen as a Bayesian inference problem, where the posterior distribution of the PDM parameters (and pose) can be inferred in a maximum a posteriori (MAP) sense. We present a novel Bayesian global optimization strategy, where the prior is used to encode the dynamic transitions of the PDM parameters. Using recursive Bayesian estimation we model the prior distribution of the data as being Gaussian. The mean and covariance were assumed to be unknown and treated as random variables. The Shape Model: The shape of a PDM is represented by the 2D locations of a mesh s = (x1,y1, . . . ,xv,yv) (v landmarks). Applying PCA on training examples, results in the parametric model s = s0 + Φb + Ψq, where s0 is the mean shape, Φ is the shape subspace matrix (n eigenvectors), b is a vector of shape parameters, q the pose parameters vector and Ψ holds four special eigenvectors that linearly model the 2D pose [4]. Goal: Given a 2v vector of observed positions y, the goal is to find the optimal set of parameters b that maximizes the posterior probability of being aligned. Using an Bayesian approach, the shape parameters are

Background modelling on tensor field has recently been proposed for foreground detection tasks. Taking into account the Riemannian structure of the tensor manifold, recent research has focused on developing parametric methods on the tensor domain e.g. gaussians mixtures (GMM) [7]. However, in some scenarios, simple parametric models do not accurately explain the physical processes. Kernel density estimators (KDE) have been successful to model, on Euclidean sample spaces the nonparametric nature of complex, time varying, and non-static backgrounds [8]. Founded on the mathematically rigorous KDE paradigm on general Riemannian manifolds [15], we define a KDE specifically to operate on the tensor manifold. We present a mathematically-sound framework for nonparametric modeling on tensor field to foreground segmentation. We endow the tensor manifold with two well-founded Riemannian metrics, i.e. Affine-Invariant and Log-Euclidean. Theoretical aspects are defined and the metrics are compared experimentally. Theoretic analysis and experimental results demonstrate the promise/effectiveness of the framework.

Multiple object tracking has been formulated recently as a global optimization problem, and solved efficiently with optimal methods such as the Hungarian Algorithm. A severe limitation is the inability to model multiple objects that are merged into a single measurement, and track them as a group, while retaining optimality. This work presents a new graph structure that encodes these multiple-match events as standard one-to-one matches, allowing computation of the solution in polynomial time. Since identities are lost when objects merge, an efficient method to identify groups is also presented, as a flow circulation problem. The problem of tracking individual objects across groups is then posed as a standard optimal assignment. Experiments show increased performance on the PETS 2006 and 2009 datasets compared to state-of-the-art algorithms.

In this paper is proposed a novel statistical learning approach, to identify cast shadows, and model their generation. We exploit the theoretically well-founded directional statistics field, in order to formulate the generation of cast shadows as a Mixture of Von Mises-Fisher distributions (MovMF) on the unit sphere. This formulation is based on a bi-illuminant physical model of cast shadows, where no prior assumptions of the spectral power distribution (SPD) of the direct light sources and ambient illumination in the scene are made. Founded on a rigorous directional statistics approach, this parametric framework is capable of modelling the shaded surface behavior in complex illumination scenes and meet real time requirements. This better model discriminating cast shadows provides a more compact representation, and achieve better accuracy, with less data and much less computation time, compared with non-parametric models previously proposed. Theoretic analysis and experimental evaluations demonstrate the effectiveness of the proposed framework.

Matching methods such as the Hungarian algorithm have recently made an appearance as an alternative to classical tracking algorithms in computer vision, since they are able to find the set of tracks that optimizes well-defined criteria over a given video sequence. However, despite being globally optimal, they carry a cost: since they require complete knowledge of the sequence, such methods cannot work with continuous video streams, a crucial requirement of realistic video surveillance applications. We were able to use the recently proposed Dynamic Hungarian Algorithm in an innovative way, adapting it to the well-known sliding window methodology. The algorithm is able to run in real-time, while retaining its optimality. We tested our implementation on several datasets, tracking humans and vehicles, and obtained reliable results using the same set of parameters on all sequences.

Statistical modeling in color space is a widely used approach for background modeling to foreground segmentation. Nevertheless, sometimes computing such statistics directly on image values is not enough to achieve a good discrimination. Thus the image may be converted into a more information rich form, such as a tensor field, in which can be encoded color and gradients. In this paper, we exploit the theoretically well-founded differential geometrical properties of the Riemannian manifold where tensors lie. We propose a novel and efficient approach for foreground segmentation on tensor field based on data modeling by means of Gaussians mixtures (GMM) directly in the tensor domain. We introduced a Expectation Maximization (EM) algorithm to estimate the mixture parameters, and are proposed two algorithms based on an online K-means approximation of EM, in order to speed up the process. Theoretic analysis and experimental evaluations demonstrate the promise and effectiveness of the proposed framework.