We recover the underlying 3D structure from images of cartoons and anime depicting the same scene. This is an interesting problem domain because images in creative media are often depicted without explicit geometric consistency for storytelling and creative expression-they are only 3D in a qualitative sense. While humans can easily perceive the underlying 3D scene from these images, existing Structure-from-Motion (SfM) methods that assume 3D consistency fail catastrophically. We present Toon3D for reconstructing geometrically inconsistent images. Our key insight is to deform the input images while recovering camera poses and scene geometry, effectively explaining away geometrical inconsistencies to achieve consistency. This process is guided by the structure inferred from monocular depth predictions. We curate a dataset with multi-view imagery from cartoons and anime that we annotate with reliable sparse correspondences using our user-friendly annotation tool. Our recovered point clouds can be plugged into novel-view synthesis methods to experience cartoons from viewpoints never drawn before. We evaluate against classical and recent learning-based SfM methods, where Toon3D is able to obtain more reliable camera poses and scene geometry.

We show that the GPS tags contained in photo metadata provide a useful control signal for image generation. We train GPS-to-image models and use them for tasks that require a fine-grained understanding of how images vary within a city. In particular, we train a diffusion model to generate images conditioned on both GPS and text. The learned model generates images that capture the distinctive appearance of different neighborhoods, parks, and landmarks. We also extract 3D models from 2D GPS-to-image models through score distillation sampling, using GPS conditioning to constrain the appearance of the reconstruction from each viewpoint. Our evaluations suggest that our GPS-conditioned models successfully learn to generate images that vary based on location, and that GPS conditioning improves estimated 3D structure.

We present a unified framework capable of solving a broad range of 3D tasks. Our approach features a stateful recurrent model that continuously updates its state representation with each new observation. Given a stream of images, this evolving state can be used to generate metric-scale pointmaps (per-pixel 3D points) for each new input in an online fashion. These pointmaps reside within a common coordinate system, and can be accumulated into a coherent, dense scene reconstruction that updates as new images arrive. Our model, called CUT3R (Continuous Updating Transformer for 3D Reconstruction), captures rich priors of real-world scenes: not only can it predict accurate pointmaps from image observations, but it can also infer unseen regions of the scene by probing at virtual, unobserved views. Our method is simple yet highly flexible, naturally accepting varying length of images that may be either video streams or unordered photo collections, containing both static and dynamic content. We evaluate our method on various 3D/4D tasks and demonstrate competitive or state-of-the-art performance in each. Project page: https://cut3r.github.io/.

We interpret the function of individual neurons in CLIP by automatically describing them using text. Analyzing the direct effects (i.e. the flow from a neuron through the residual stream to the output) or the indirect effects (overall contribution) fails to capture the neurons' function in CLIP. Therefore, we present the "second-order lens", analyzing the effect flowing from a neuron through the later attention heads, directly to the output. We find that these effects are highly selective: for each neuron, the effect is significant for <2% of the images. Moreover, each effect can be approximated by a single direction in the text-image space of CLIP. We describe neurons by decomposing these directions into sparse sets of text representations. The sets reveal polysemantic behavior - each neuron corresponds to multiple, often unrelated, concepts (e.g. ships and cars). Exploiting this neuron polysemy, we mass-produce "semantic" adversarial examples by generating images with concepts spuriously correlated to the incorrect class. Additionally, we use the second-order effects for zero-shot segmentation, outperforming previous methods. Our results indicate that a automated interpretation of neurons can be used for model deception and for introducing new model capabilities

Sample-efficient online reinforcement learning often uses replay buffers to store experience for reuse when updating the value function. However, uniform replay is inefficient, since certain classes of transitions can be more relevant to learning. While prioritization of more useful samples is helpful, this strategy can also lead to overfitting, as useful samples are likely to be more rare. In this work, we instead propose a prioritized, parametric version of an agent's memory, using generative models to capture online experience. This paradigm enables (1) densification of past experience, with new generations that benefit from the generative model's generalization capacity and (2) guidance via a family of "relevance functions" that push these generations towards more useful parts of an agent's acquired history. We show this recipe can be instantiated using conditional diffusion models and simple relevance functions such as curiosity- or value-based metrics. Our approach consistently improves performance and sample efficiency in both state- and pixel-based domains. We expose the mechanisms underlying these gains, showing how guidance promotes diversity in our generated transitions and reduces overfitting. We also showcase how our approach can train policies with even higher update-to-data ratios than before, opening up avenues to better scale online RL agents. Project page available at: https://pgenreplay.github.io

Deep learning models often struggle when deployed in real-world settings due to distribution shifts between training and test data. While existing approaches like domain adaptation and test-time training (TTT) offer partial solutions, they typically require additional data or domain-specific auxiliary tasks. We present Idempotent Test-Time Training (IT3), a novel approach that enables on-the-fly adaptation to distribution shifts using only the current test instance, without any auxiliary task design. Our key insight is that enforcing idempotence---where repeated applications of a function yield the same result---can effectively replace domain-specific auxiliary tasks used in previous TTT methods. We theoretically connect idempotence to prediction confidence and demonstrate that minimizing the distance between successive applications of our model during inference leads to improved out-of-distribution performance. Extensive experiments across diverse domains (including image classification, aerodynamics prediction, and aerial segmentation) and architectures (MLPs, CNNs, GNNs) show that IT3 consistently outperforms existing approaches while being simpler and more widely applicable. Our results suggest that idempotence provides a universal principle for test-time adaptation that generalizes across domains and architectures.

This paper proposes a novel scene understanding task called Visual Jenga. Drawing inspiration from the game Jenga, the proposed task involves progressively removing objects from a single image until only the background remains. Just as Jenga players must understand structural dependencies to maintain tower stability, our task reveals the intrinsic relationships between scene elements by systematically exploring which objects can be removed while preserving scene coherence in both physical and geometric sense. As a starting point for tackling the Visual Jenga task, we propose a simple, data-driven, training-free approach that is surprisingly effective on a range of real-world images. The principle behind our approach is to utilize the asymmetry in the pairwise relationships between objects within a scene and employ a large inpainting model to generate a set of counterfactuals to quantify the asymmetry.

We investigate the mechanism underlying a previously identified phenomenon in Vision Transformers -- the emergence of high-norm tokens that lead to noisy attention maps (Darcet et al., 2024). We observe that in multiple models (e.g., CLIP, DINOv2), a sparse set of neurons is responsible for concentrating high-norm activations on outlier tokens, leading to irregular attention patterns and degrading downstream visual processing. While the existing solution for removing these outliers involves retraining models from scratch with additional learned $\textit{register tokens}$, we use our findings to create a training-free approach to mitigate these artifacts. By shifting the high-norm activations from our discovered $\textit{register neurons}$ into an additional untrained token, we can mimic the effect of register tokens on a model already trained without registers. We demonstrate that our method produces cleaner attention and feature maps, enhances performance over base models across multiple downstream visual tasks, and achieves results comparable to models explicitly trained with register tokens. We then extend test-time registers to off-the-shelf vision-language models, yielding cleaner attention-based, text-to-image attribution. Finally, we outline a simple mathematical model that reflects the observed behavior of register neurons and high norm tokens. Our results suggest that test-time registers effectively take on the role of register tokens at test-time, offering a training-free solution for any pre-trained model released without them.

Data attribution for text-to-image models aims to identify the training images that most significantly influenced a generated output. Existing attribution methods involve considerable computational resources for each query, making them impractical for real-world applications. We propose a novel approach for scalable and efficient data attribution. Our key idea is to distill a slow, unlearning-based attribution method to a feature embedding space for efficient retrieval of highly influential training images. During deployment, combined with efficient indexing and search methods, our method successfully finds highly influential images without running expensive attribution algorithms. We show extensive results on both medium-scale models trained on MSCOCO and large-scale Stable Diffusion models trained on LAION, demonstrating that our method can achieve better or competitive performance in a few seconds, faster than existing methods by 2,500x - 400,000x. Our work represents a meaningful step towards the large-scale application of data attribution methods on real-world models such as Stable Diffusion.

Prior work has established Test-Time Training (TTT) as a general framework to further improve a trained model at test time. Before making a prediction on each test instance, the model is first trained on the same instance using a self-supervised task such as reconstruction. We extend TTT to the streaming setting, where multiple test instances - video frames in our case - arrive in temporal order. Our extension is online TTT: The current model is initialized from the previous model, then trained on the current frame and a small window of frames immediately before. Online TTT significantly outperforms the fixed-model baseline for four tasks, on three real-world datasets. The improvements are more than 2.2x and 1.5x for instance and panoptic segmentation. Surprisingly, online TTT also outperforms its offline variant that accesses strictly more information, training on all frames from the entire test video regardless of temporal order. This finding challenges those in prior work using synthetic videos. We formalize a notion of locality as the advantage of online over offline TTT, and analyze its role with ablations and a theory based on bias-variance trade-off.

We introduce a novel sequential modeling approach which enables learning a Large Vision Model (LVM) without making use of any linguistic data. To do this, we define a common format, “visual sentences ”, in which we can represent raw images and videos as well as annotated data sources such as semantic segmentations and depth reconstructions with-out needing any meta-knowledge beyond the pixels. Once this wide variety of visual data (comprising 420 billion to-kens) is represented as sequences, the model can be trained to minimize a cross-entropy loss for next token prediction. By training across various scales of model architecture and data diversity, we provide empirical evidence that our models scale effectively. Many different vision tasks can be solved by designing suitable visual prompts at test time.

While neural rendering has led to impressive advances in scene reconstruction and novel view synthesis, it relies heavily on accurately pre-computed camera poses. To relax this constraint, multiple efforts have been made to train Neural Radiance Fields (NeRFs) without pre-processed camera poses. However, the implicit representations of NeRFs provide extra challenges to optimize the 3D structure and camera poses at the same time. On the other hand, the recently proposed 3D Gaussian Splatting provides new opportunities given its explicit point cloud representations. This paper leverages both the explicit geometric representation and the continuity of the input video stream to perform novel view synthesis without any SfM preprocessing. We process the input frames in a sequential manner and progressively grow the 3D Gaussians set by taking one input frame at a time, without the need to pre-compute the camera poses. Our method significantly improves over previous approaches in view synthesis and camera pose estimation under large motion changes. Our project page is https://oasisyang.github.io/colmap-free-3dgs.

We investigate the CLIP image encoder by analyzing how individual model components affect the final representation. We decompose the image representation as a sum across individual image patches, model layers, and attention heads, and use CLIP's text representation to interpret the summands. Interpreting the attention heads, we characterize each head's role by automatically finding text representations that span its output space, which reveals property-specific roles for many heads (e.g. location or shape). Next, interpreting the image patches, we uncover an emergent spatial localization within CLIP. Finally, we use this understanding to remove spurious features from CLIP and to create a strong zero-shot image segmenter. Our results indicate that scalable understanding of transformer models is attainable and can be used to repair and improve models.

We propose a new approach for generative modeling based on training a neural network to be idempotent. An idempotent operator is one that can be applied sequentially without changing the result beyond the initial application, namely $f(f(z))=f(z)$. The proposed model $f$ is trained to map a source distribution (e.g, Gaussian noise) to a target distribution (e.g. realistic images) using the following objectives: (1) Instances from the target distribution should map to themselves, namely $f(x)=x$. We define the target manifold as the set of all instances that $f$ maps to themselves. (2) Instances that form the source distribution should map onto the defined target manifold. This is achieved by optimizing the idempotence term, $f(f(z))=f(z)$ which encourages the range of $f(z)$ to be on the target manifold. Under ideal assumptions such a process provably converges to the target distribution. This strategy results in a model capable of generating an output in one step, maintaining a consistent latent space, while also allowing sequential applications for refinement. Additionally, we find that by processing inputs from both target and source distributions, the model adeptly projects corrupted or modified data back to the target manifold. This work is a first step towards a ``global projector'' that enables projecting any input into a target data distribution.

We introduce a method to generate 3D scenes that are disentangled into their component objects. This disentanglement is unsupervised, relying only on the knowledge of a large pretrained text-to-image model. Our key insight is that objects can be discovered by finding parts of a 3D scene that, when rearranged spatially, still produce valid configurations of the same scene. Concretely, our method jointly optimizes multiple NeRFs---each representing its own object---along with a *set of layouts* that composite these objects into scenes. We then encourage these composited scenes to be in-distribution according to the image generator. We show that despite its simplicity, our approach successfully generates 3D scenes decomposed into individual objects, enabling new capabilities in text-to-3D content creation.

We introduce a benchmark to directly evaluate the alignment between human observers and vision models on a 3D shape inference task. We leverage an experimental design from the cognitive sciences: given a set of images, participants identify which contain the same/different objects, despite considerable viewpoint variation. We draw from a diverse range of images that include common objects (e.g., chairs) as well as abstract shapes (i.e., procedurally generated 'nonsense' objects). After constructing over 2000 unique image sets, we administer these tasks to human participants, collecting 35K trials of behavioral data from over 500 participants. This includes explicit choice behaviors as well as intermediate measures, such as reaction time and gaze data. We then evaluate the performance of common vision models (e.g., DINOv2, MAE, CLIP). We find that humans outperform all models by a wide margin. Using a multi-scale evaluation approach, we identify underlying similarities and differences between models and humans: while human-model performance is correlated, humans allocate more time/processing on challenging trials. All images, data, and code can be accessed via our project page.

We investigate the space of weights spanned by a large collection of customized diffusion models. We populate this space by creating a dataset of over 60,000 models, each of which is a base model fine-tuned to insert a different person's visual identity. We model the underlying manifold of these weights as a subspace, which we term $\textit{weights2weights}$. We demonstrate three immediate applications of this space that result in new diffusion models -- sampling, editing, and inversion. First, sampling a set of weights from this space results in a new model encoding a novel identity. Next, we find linear directions in this space corresponding to semantic edits of the identity (e.g., adding a beard), resulting in a new model with the original identity edited. Finally, we show that inverting a single image into this space encodes a realistic identity into a model, even if the input image is out of distribution (e.g., a painting). We further find that these linear properties of the diffusion model weight space extend to other visual concepts. Our results indicate that the weight space of fine-tuned diffusion models can behave as an interpretable $\textit{meta}$-latent space producing new models.

Score distillation sampling (SDS) has proven to be an important tool, enabling the use of large-scale diffusion priors for tasks operating in data-poor domains. Unfortunately, SDS has a number of characteristic artifacts that limit its utility in general-purpose applications. In this paper, we make progress toward understanding the behavior of SDS and its variants by viewing them as solving an optimal-cost transport path from some current source distribution to a target distribution. Under this new interpretation, we argue that these methods' characteristic artifacts are caused by (1) linear approximation of the optimal path and (2) poor estimates of the source distribution. We show that by calibrating the text conditioning of the source distribution, we can produce high-quality generation and translation results with little extra overhead. Our method can be easily applied across many domains, matching or beating the performance of specialized methods. We demonstrate its utility in text-to-2D, text-to-3D, translating paintings to real images, optical illusion generation, and 3D sketch-to-real. We compare our method to existing approaches for score distillation sampling and show that it can produce high-frequency details with realistic colors.

The goal of data attribution for text-to-image models is to identify the training images that most influence the generation of a new image. Influence is defined such that, for a given output, if a model is retrained from scratch without the most influential images, the model would fail to reproduce the same output. Unfortunately, directly searching for these influential images is computationally infeasible, since it would require repeatedly retraining models from scratch. In our work, we propose an efficient data attribution method by simulating unlearning the synthesized image. We achieve this by increasing the training loss on the output image, without catastrophic forgetting of other, unrelated concepts. We then identify training images with significant loss deviations after the unlearning process and label these as influential. We evaluate our method with a computationally intensive but "gold-standard" retraining from scratch and demonstrate our method's advantages over previous methods.

We propose a method for editing images from human instructions: given an input image and a written instruction that tells the model what to do, our model follows these instructions to edit the image. To obtain training data for this problem, we combine the knowledge of two large pretrained models—a language model (GPT-3) and a text-to-image model (Stable Diffusion)—to generate a large dataset of image editing examples. Our conditional diffusion model, InstructPix2Pix, is trained on our generated data, and generalizes to real images and user-written instructions at inference time. Since it performs edits in the forward pass and does not require per-example fine-tuning or inversion, our model edits images quickly, in a matter of seconds. We show compelling editing results for a diverse collection of input images and written instructions.

Dataset Distillation aims to distill an entire dataset's knowledge into a few synthetic images. The idea is to synthe-size a small number of synthetic data points that, when given to a learning algorithm as training data, result in a model approximating one trained on the original data. Despite a recent upsurge of progress in the field, existing dataset dis-tillation methods fail to generalize to new architectures and scale to high-resolution datasets. To overcome the above issues, we propose to use the learned prior from pretrained deep generative models to synthesize the distilled data. To achieve this, we present a new optimization algorithm that distills a large number of images into a few intermediate feature vectors in the generative model's latent space. Our method augments existing techniques, significantly improving cross-architecture generalization in all settings.

We study the problem of inferring scene affordances by presenting a method for realistically inserting people into scenes. Given a scene image with a marked region and an image of a person, we insert the person into the scene while respecting the scene affordances. Our model can infer the set of realistic poses given the scene context, re-pose the reference person, and harmonize the composition. We set up the task in a self-supervised fashion by learning to repose humans in video clips. We train a large-scale diffusion model on a dataset of 2.4M video clips that produces diverse plausible poses while respecting the scene context. Given the learned human-scene composition, our model can also hallucinate realistic people and scenes when prompted without conditioning and also enables interactive editing. A quantitative evaluation shows that our method synthesizes more realistic human appearance and more natural human-scene interactions than prior work.

Do different neural networks, trained for various vision tasks, share some common representations? In this paper, we demonstrate the existence of common features we call "Rosetta Neurons" across a range of models with different architectures, different tasks (generative and discriminative), and different types of supervision (class-supervised, text-supervised, self-supervised). We present an algorithm for mining a dictionary of Rosetta Neurons across several popular vision models: Class Supervised-ResNet50, DINO-ResNet50, DINO-ViT, MAE, CLIP-ResNet50, Big-GAN, StyleGAN-2, StyleGAN-XL. Our findings suggest that certain visual concepts and structures are inherently embedded in the natural world and can be learned by different models regardless of the specific task or architecture, and without the use of semantic labels. We can visualize shared concepts directly due to generative models included in our analysis. The Rosetta Neurons facilitate model-to-model translation enabling various inversion-based manipulations, including cross-class alignments, shifting, zooming, and more, without the need for specialized training.

We propose a method for editing NeRF scenes with text-instructions. Given a NeRF of a scene and the collection of images used to reconstruct it, our method uses an image-conditioned diffusion model (InstructPix2Pix) to iteratively edit the input images while optimizing the underlying scene, resulting in an optimized 3D scene that respects the edit instruction. We demonstrate that our proposed method is able to edit large-scale, real-world scenes, and is able to accomplish more realistic, targeted edits than prior work. Result videos can be found on the project website: https://instruct-nerf2nerf.github.io.

While large text-to-image models are able to synthesize "novel" images, these images are necessarily a reflection of the training data. The problem of data attribution in such models – which of the images in the training set are most responsible for the appearance of a given generated image – is a difficult yet important one. As an initial step toward this problem, we evaluate attribution through "customization" methods, which tune an existing large-scale model toward a given exemplar object or style. Our key insight is that this allow us to efficiently create synthetic images that are computationally influenced by the exemplar by construction. With our new dataset of such exemplar-influenced images, we are able to evaluate various data attribution algorithms and different possible feature spaces. Furthermore, by training on our dataset, we can tune standard models, such as DINO, CLIP, and ViT, toward the attribution problem. Even though the procedure is tuned towards small exemplar sets, we show generalization to larger sets. Finally, by taking into account the inherent uncertainty of the problem, we can assign soft attribution scores over a set of training images.

Vision models typically rely on fine-tuning general-purpose models pre-trained on large, static datasets. These general-purpose models only capture the knowledge within their pre-training datasets, which are tiny, out-of-date snapshots of the Internet—where billions of images are uploaded each day. We suggest an alternate approach: rather than hoping our static datasets transfer to our desired tasks after large-scale pre-training, we propose dynamically utilizing the Internet to quickly train a small-scale model that does extremely well on a target dataset. Our approach, called Internet Explorer, explores the web in a self-supervised manner to progressively find relevant examples that improve performance on a desired target dataset. It cycles between searching for images on the Internet with text queries, self-supervised training on downloaded images, determining which images were useful, and prioritizing what to search for next. We evaluate Internet Explorer across several datasets and show that it outperforms or matches CLIP oracle performance using just a single GPU desktop to actively query the Internet for 30-40 hours.

Large-scale generative models are capable of producing high-quality images from detailed prompts. However, many aspects of an image are difficult or impossible to convey through text. We introduce self-guidance, a method that provides precise control over properties of the generated image by guiding the internal representations of diffusion models. We demonstrate that the size, location, and appearance of objects can be extracted from these representations, and show how to use them to steer the sampling process. Self-guidance operates similarly to standard classifier guidance, but uses signals present in the pretrained model itself, requiring no additional models or training. We demonstrate the flexibility and effectiveness of self-guided generation through a wide range of challenging image manipulations, such as modifying the position or size of a single object (keeping the rest of the image unchanged), merging the appearance of objects in one image with the layout of another, composing objects from multiple images into one, and more. We also propose a new method for reconstruction using self-guidance, which allows extending our approach to editing real images.

Given a set of calibrated images of a scene, we present an approach that produces a simple, compact, and actionable 3D world representation by means of 3D primitives. While many approaches focus on recovering high-fidelity 3D scenes, we focus on parsing a scene into mid-level 3D representations made of a small set of textured primitives. Such representations are interpretable, easy to manipulate and suited for physics-based simulations. Moreover, unlike existing primitive decomposition methods that rely on 3D input data, our approach operates directly on images through differentiable rendering. Specifically, we model primitives as textured superquadric meshes and optimize their parameters from scratch with an image rendering loss. We highlight the importance of modeling transparency for each primitive, which is critical for optimization and also enables handling varying numbers of primitives. We show that the resulting textured primitives faithfully reconstruct the input images and accurately model the visible 3D points, while providing amodal shape completions of unseen object regions. We compare our approach to the state of the art on diverse scenes from DTU, and demonstrate its robustness on real-life captures from BlendedMVS and Nerfstudio. We also showcase how our results can be used to effortlessly edit a scene or perform physical simulations. Code and video results are available at https://www.tmonnier.com/DBW.

A range of video modeling tasks, from optical flow to multiple object tracking, share the same fundamental challenge: establishing space-time correspondence. Yet, approaches that dominate each space differ. We take a step to-wards bridging this gap by extending the recent contrastive random walk formulation to much denser, pixel-level spacetime graphs. The main contribution is introducing hierarchy into the search problem by computing the transition matrix between two frames in a coarse-to-fine manner, forming a multiscale contrastive random walk when ex-tended in time. This establishes a unified technique for self-supervised learning of optical flow, keypoint tracking, and video object segmentation. Experiments demonstrate that, for each of these tasks, the unified model achieves performance competitive with strong self-supervised approaches specific to that task.11Project page at https://jasonbian97.github.io/flowwalk

Dataset distillation is the task of synthesizing a small dataset such that a model trained on the synthetic set will match the test accuracy of the model trained on the full dataset. In this paper, we propose a new formulation that optimizes our distilled data to guide networks to a similar state as those trained on real data across many training steps. Given a network, we train it for several iterations on our distilled data and optimize the distilled data with respect to the distance between the synthetically trained parameters and the parameters trained on real data. To efficiently obtain the initial and target network parameters for large-scale datasets, we pre-compute and store training trajectories of expert networks trained on the real dataset. Our method handily outperforms existing methods and also allows us to distill higher-resolution visual data.

We propose GAN-Supervised Learning, a framework for learning discriminative models and their GAN-generated training data jointly end-to-end. We apply our framework to the dense visual alignment problem. Inspired by the classic Congealing method, our GAN gealing algorithm trains a Spatial Transformer to map random samples from a GAN trained on unaligned data to a common, jointly-learned target mode. We show results on eight datasets, all of which demonstrate our method successfully aligns complex data and discovers dense correspondences. GANgealing significantly outperforms past self-supervised correspondence algorithms and performs on-par with (and sometimes exceeds) state-of-the-art supervised correspondence algorithms on several datasets-without making use of any correspondence supervision or data augmentation and despite being trained exclusively on GAN-generated data. For precise correspondence, we improve upon state-of-the-art supervised methods by as much as 3 ×. We show applications of our method for augmented reality, image editing and automated pre-processing of image datasets for downstream GAN training.

How does one adapt a pre-trained visual model to novel downstream tasks without task-specific finetuning or any model modification? Inspired by prompting in NLP, this paper investigates visual prompting: given input-output image example(s) of a new task at test time and a new input image, the goal is to automatically produce the output image, consistent with the given examples. We show that posing this problem as simple image inpainting -- literally just filling in a hole in a concatenated visual prompt image -- turns out to be surprisingly effective, provided that the inpainting algorithm has been trained on the right data. We train masked auto-encoders on a new dataset that we curated -- 88k unlabeled figures from academic papers sources on Arxiv. We apply visual prompting to these pretrained models and demonstrate results on various downstream image-to-image tasks, including foreground segmentation, single object detection, colorization, edge detection, etc. Project page: https://yossigandelsman.github.io/visual_prompt

We present a video generation model that accurately reproduces object motion, changes in camera viewpoint, and new content that arises over time. Existing video generation methods often fail to produce new content as a function of time while maintaining consistencies expected in real environments, such as plausible dynamics and object persistence. A common failure case is for content to never change due to over-reliance on inductive bias to provide temporal consistency, such as a single latent code that dictates content for the entire video. On the other extreme, without long-term consistency, generated videos may morph unrealistically between different scenes. To address these limitations, we prioritize the time axis by redesigning the temporal latent representation and learning long-term consistency from data by training on longer videos. We leverage a two-phase training strategy, where we separately train using longer videos at a low resolution and shorter videos at a high resolution. To evaluate the capabilities of our model, we introduce two new benchmark datasets with explicit focus on long-term temporal dynamics.

Test-time training adapts to a new test distribution on the fly by optimizing a model for each test input using self-supervision.In this paper, we use masked autoencoders for this one-sample learning problem.Empirically, our simple method improves generalization on many visual benchmarks for distribution shifts.Theoretically, we characterize this improvement in terms of the bias-variance trade-off.

We introduce a non-parametric approach for infinite video texture synthesis using a representation learned via contrastive learning. We take inspiration from Video Textures [43], which showed that plausible new videos could be generated from a single one by stitching its frames together in a novel yet consistent order. This classic work, however, was constrained by its use of hand-designed distance metrics, limiting its use to simple, repetitive videos. We draw on recent techniques from self-supervised learning to learn this distance metric, allowing us to compare frames in a manner that scales to more challenging dynamics, and to condition on other data, such as audio. We learn representations for video frames and frame-to-frame transition probabilities by fitting a video-specific model trained using contrastive learning. To synthesize a texture, we randomly sample frames with high transition probabilities to generate diverse temporally smooth videos with novel sequences and transitions. The model naturally extends to an audio-conditioned setting without requiring any finetuning. Our model outperforms baselines on human perceptual scores, can handle a diverse range of input videos, and can combine semantic and audiovisual cues in order to synthesize videos that synchronize well with an audio signal.

Training generative models, such as GANs, on a target domain containing limited examples (e.g., 10) can easily result in overfitting. In this work, we seek to utilize a large source domain for pretraining and transfer the diversity information from source to target. We propose to preserve the relative similarities and differences between instances in the source via a novel cross-domain distance consistency loss. To further reduce overfitting, we present an anchor-based strategy to encourage different levels of realism over different regions in the latent space. With extensive results in both photorealistic and non-photorealistic domains, we demonstrate qualitatively and quantitatively that our few-shot model automatically discovers correspondences between source and target domains and generates more diverse and realistic images than previous methods.

A video autoencoder is proposed for learning disentangled representations of 3D structure and camera pose from videos in a self-supervised manner. Relying on temporal continuity in videos, our work assumes that the 3D scene structure in nearby video frames remains static. Given a sequence of video frames as input, the video autoencoder extracts a disentangled representation of the scene including: (i) a temporally-consistent deep voxel feature to represent the 3D structure and (ii) a 3D trajectory of camera pose for each frame. These two representations will then be re-entangled for rendering the input video frames. This video autoencoder can be trained directly using a pixel re-construction loss, without any ground truth 3D or camera pose annotations. The disentangled representation can be applied to a range of tasks, including novel view synthesis, camera pose estimation, and video generation by motion following. We evaluate our method on several large-scale natural video datasets, and show generalization results on out-of-domain images. Project page with code: https://zlai0.github.io/VideoAutoencoder.

Artists and video game designers often construct 2D animations using libraries of sprites---textured patches of objects and characters. We propose a deep learning approach that decomposes sprite-based video animations into a disentangled representation of recurring graphic elements in a self-supervised manner. By jointly learning a dictionary of possibly transparent patches and training a network that places them onto a canvas, we deconstruct sprite-based content into a sparse, consistent, and explicit representation that can be easily used in downstream tasks, like editing or analysis. Our framework offers a promising approach for discovering recurring visual patterns in image collections without supervision.

In this work we ask whether it is possible to create a "universal" detector for telling apart real images from these generated by a CNN, regardless of architecture or dataset used. To test this, we collect a dataset consisting of fake images generated by 11 different CNN-based image generator models, chosen to span the space of commonly used architectures today (ProGAN, StyleGAN, BigGAN, CycleGAN, StarGAN, GauGAN, DeepFakes, cascaded refinement networks, implicit maximum likelihood estimation, second-order attention super-resolution, seeing-in-the-dark). We demonstrate that, with careful pre- and post-processing and data augmentation, a standard image classifier trained on only one specific CNN generator (ProGAN) is able to generalize surprisingly well to unseen architectures, datasets, and training methods (including the just released StyleGAN2). Our findings suggest the intriguing possibility that today's CNN-generated images share some common systematic flaws, preventing them from achieving realistic image synthesis.

In this paper, we propose Test-Time Training, a general approach for improving the performance of predictive models when training and test data come from different distributions. We turn a single unlabeled test sample into a self-supervised learning problem, on which we update the model parameters before making a prediction. This also extends naturally to data in an online stream. Our simple approach leads to improvements on diverse image classification benchmarks aimed at evaluating robustness to distribution shifts.

This paper proposes a simple self-supervised approach for learning a representation for visual correspondence from raw video. We cast correspondence as prediction of links in a space-time graph constructed from video. In this graph, the nodes are patches sampled from each frame, and nodes adjacent in time can share a directed edge. We learn a representation in which pairwise similarity defines transition probability of a random walk, such that prediction of long-range correspondence is computed as a walk along the graph. We optimize the representation to place high probability along paths of similarity. Targets for learning are formed without supervision, by cycle-consistency: the objective is to maximize the likelihood of returning to the initial node when walking along a graph constructed from a palindrome of frames. Thus, a single path-level constraint implicitly supervises chains of intermediate comparisons. When used as a similarity metric without adaptation, the learned representation outperforms the self-supervised state-of-the-art on label propagation tasks involving objects, semantic parts, and pose. Moreover, we demonstrate that a technique we call edge dropout, as well as self-supervised adaptation at test-time, further improve transfer for object-centric correspondence.

Deep generative models have become increasingly effective at producing realistic images from randomly sampled seeds, but using such models for controllable manipulation of existing images remains challenging. We propose the Swapping Autoencoder, a deep model designed specifically for image manipulation, rather than random sampling. The key idea is to encode an image into two independent components and enforce that any swapped combination maps to a realistic image. In particular, we encourage the components to represent structure and texture, by enforcing one component to encode co-occurrent patch statistics across different parts of the image. As our method is trained with an encoder, finding the latent codes for a new input image becomes trivial, rather than cumbersome. As a result, our method enables us to manipulate real input images in various ways, including texture swapping, local and global editing, and latent code vector arithmetic. Experiments on multiple datasets show that our model produces better results and is substantially more efficient compared to recent generative models.

Our goal in this paper is to discover near duplicate patterns in large collections of artworks. This is harder than standard instance mining due to differences in the artistic media (oil, pastel, drawing, etc), and imperfections inherent in the copying process. Our key technical insight is to adapt a standard deep feature to this task by fine-tuning it on the specific art collection using self-supervised learning. More specifically, spatial consistency between neighbouring feature matches is used as supervisory fine-tuning signal. The adapted feature leads to more accurate style invariant matching, and can be used with a standard discovery approach, based on geometric verification, to identify duplicate patterns in the dataset. The approach is evaluated on several different datasets and shows surprisingly good qualitative discovery results. For quantitative evaluation of the method, we annotated 273 near duplicate details in a dataset of 1587 artworks attributed to Jan Brueghel and his workshop. Beyond artworks, we also demonstrate improvement on localization on the Oxford5K photo dataset as well as on historical photograph localization on the Large Time Lags Location (LTLL) dataset.

We introduce a self-supervised method for learning visual correspondence from unlabeled video. The main idea is to use cycle-consistency in time as free supervisory signal for learning visual representations from scratch. At training time, our model learns a feature map representation to be useful for performing cycle-consistent tracking. At test time, we use the acquired representation to find nearest neighbors across space and time. We demonstrate the generalizability of the representation -- without finetuning -- across a range of visual correspondence tasks, including video object segmentation, keypoint tracking, and optical flow. Our approach outperforms previous self-supervised methods and performs competitively with strongly supervised methods.

This paper presents a simple method for “do as I do” motion transfer: given a source video of a person dancing, we can transfer that performance to a novel (amateur) target after only a few minutes of the target subject performing standard moves. We approach this problem as video-to-video translation using pose as an intermediate representation. To transfer the motion, we extract poses from the source subject and apply the learned pose-to-appearance mapping to generate the target subject. We predict two consecutive frames for temporally coherent video results and introduce a separate pipeline for realistic face synthesis. Although our method is quite simple, it produces surprisingly compelling results (see video). This motivates us to also provide a forensics tool for reliable synthetic content detection, which is able to distinguish videos synthesized by our system from real data. In addition, we release a first-of-its-kind open-source dataset of videos that can be legally used for training and motion transfer.

We propose an interactive GAN-based sketch-to-image translation method that helps novice users easily create images of simple objects. The user starts with a sparse sketch and a desired object category, and the network then recommends its plausible completion(s) and shows a corresponding synthesized image. This enables a feedback loop, where the user can edit the sketch based on the network's recommendations, while the network is able to better synthesize the image that the user might have in mind. In order to use a single model for a wide array of object classes, we introduce a gating-based approach for class conditioning, which allows us to generate distinct classes without feature mixing, from a single generator network.

Most malicious photo manipulations are created using standard image editing tools, such as Adobe Photoshop. We present a method for detecting one very popular Photoshop manipulation -- image warping applied to human faces -- using a model trained entirely using fake images that were automatically generated by scripting Photoshop itself. We show that our model outperforms humans at the task of recognizing manipulated images, can predict the specific location of edits, and in some cases can be used to "undo" a manipulation to reconstruct the original, unedited image. We demonstrate that the system can be successfully applied to artist-created image manipulations.

Contemporary sensorimotor learning approaches typically start with an existing complex agent (e.g., a robotic arm), which they learn to control. In contrast, this paper investigates a modular co-evolution strategy: a collection of primitive agents learns to dynamically self-assemble into composite bodies while also learning to coordinate their behavior to control these bodies. Each primitive agent consists of a limb with a motor attached at one end. Limbs may choose to link up to form collectives. When a limb initiates a link-up action and there is another limb nearby, the latter is magnetically connected to the 'parent' limb's motor. This forms a new single agent, which may further link with other agents. In this way, complex morphologies can emerge, controlled by a policy whose architecture is in explicit correspondence with the morphology. We evaluate the performance of these dynamic and modular agents in simulated environments. We demonstrate better generalization to test-time changes both in the environment, as well as in the structure of the agent, compared to static and monolithic baselines. Project videos and source code are provided in the supplementary material.

We propose the first learning-based algorithm that can relight images in a plausible and controllable manner given multiple views of an outdoor scene. In particular, we introduce a geometry-aware neural network that utilizes multiple geometry cues (normal maps, specular direction, etc.) and source and target shadow masks computed from a noisy proxy geometry obtained by multi-view stereo. Our model is a three-stage pipeline: two subnetworks refine the source and target shadow masks, and a third performs the final relighting. Furthermore, we introduce a novel representation for the shadow masks, which we call RGB shadow images. They reproject the colors from all views into the shadowed pixels and enable our network to cope with inacuraccies in the proxy and the non-locality of the shadow casting interactions. Acquiring large-scale multi-view relighting datasets for real scenes is challenging, so we train our network on photorealistic synthetic data. At train time, we also compute a noisy stereo-based geometric proxy, this time from the synthetic renderings. This allows us to bridge the gap between the real and synthetic domains. Our model generalizes well to real scenes. It can alter the illumination of drone footage, image-based renderings, textured mesh reconstructions, and even internet photo collections.

A major stumbling block to progress in understanding basic human interactions, such as getting out of bed or opening a refrigerator, is lack of good training data. Most past efforts have gathered this data explicitly: starting with a laundry list of action labels, and then querying search engines for videos tagged with each label. In this work, we do the reverse and search implicitly: we start with a large collection of interaction-rich video data and then annotate and analyze it. We use Internet Lifestyle Vlogs as the source of surprisingly large and diverse interaction data. We show that by collecting the data first, we are able to achieve greater scale and far greater diversity in terms of actions and actors. Additionally, our data exposes biases built into common explicitly gathered data. We make sense of our data by analyzing the central component of interaction - hands. We benchmark two tasks: identifying semantic object contact at the video level and non-semantic contact state at the frame level. We additionally demonstrate future prediction of hands.

We present a framework for learning single-view shape and pose prediction without using direct supervision for either. Our approach allows leveraging multi-view observations from unknown poses as supervisory signal during training. Our proposed training setup enforces geometric consistency between the independently predicted shape and pose from two views of the same instance. We consequently learn to predict shape in an emergent canonical (view-agnostic) frame along with a corresponding pose predictor. We show empirical and qualitative results using the ShapeNet dataset and observe encouragingly competitive performance to previous techniques which rely on stronger forms of supervision. We also demonstrate the applicability of our framework in a realistic setting which is beyond the scope of existing techniques: using a training dataset comprised of online product images where the underlying shape and pose are unknown.

The goal of this paper is to take a single 2D image of a scene and recover the 3D structure in terms of a small set of factors: a layout representing the enclosing surfaces as well as a set of objects represented in terms of shape and pose. We propose a convolutional neural network-based approach to predict this representation and benchmark it on a large dataset of indoor scenes. Our experiments evaluate a number of practical design questions, demonstrate that we can infer this representation, and quantitatively and qualitatively demonstrate its merits compared to alternate representations.

While it is nearly effortless for humans to quickly assess the perceptual similarity between two images, the underlying processes are thought to be quite complex. Despite this, the most widely used perceptual metrics today, such as PSNR and SSIM, are simple, shallow functions, and fail to account for many nuances of human perception. Recently, the deep learning community has found that features of the VGG network trained on ImageNet classification has been remarkably useful as a training loss for image synthesis. But how perceptual are these so-called "perceptual losses"? What elements are critical for their success? To answer these questions, we introduce a new dataset of human perceptual similarity judgments. We systematically evaluate deep features across different architectures and tasks and compare them with classic metrics. We find that deep features outperform all previous metrics by large margins on our dataset. More surprisingly, this result is not restricted to ImageNet-trained VGG features, but holds across different deep architectures and levels of supervision (supervised, self-supervised, or even unsupervised). Our results suggest that perceptual similarity is an emergent property shared across deep visual representations.

What makes humans so good at solving seemingly complex video games? Unlike computers, humans bring in a great deal of prior knowledge about the world, enabling efficient decision making. This paper investigates the role of human priors for solving video games. Given a sample game, we conduct a series of ablation studies to quantify the importance of various priors on human performance. We do this by modifying the video game environment to systematically mask different types of visual information that could be used by humans as priors. We find that removal of some prior knowledge causes a drastic degradation in the speed with which human players solve the game, e.g. from 2 minutes to over 20 minutes. Furthermore, our results indicate that general priors, such as the importance of objects and visual consistency, are critical for efficient game-play. Videos and the game manipulations are available at https://rach0012.github.io/humanRL_website/

Domain adaptation is critical for success in new, unseen environments. Adversarial adaptation models have shown tremendous progress towards adapting to new environments by focusing either on discovering domain invariant representations or by mapping between unpaired image domains. While feature space methods are difficult to interpret and sometimes fail to capture pixel-level and low-level domain shifts, image space methods sometimes fail to incorporate high level semantic knowledge relevant for the end task. We propose a model which adapts between domains using both generative image space alignment and latent representation space alignment. Our approach, Cycle-Consistent Adversarial Domain Adaptation (CyCADA), guides transfer between domains according to a specific discriminatively trained task and avoids divergence by enforcing consistency of the relevant semantics before and after adaptation. We evaluate our method on a variety of visual recognition and prediction settings, including digit classification and semantic segmentation of road scenes, advancing state-of-the-art performance for unsupervised adaptation from synthetic to real world driving domains.

Light-field cameras have recently emerged as a powerful tool for one-shot passive 3D shape capture. However, obtaining the shape of glossy objects like metals or plastics remains challenging, since standard Lambertian cues like photo-consistency cannot be easily applied. In this paper, we derive a spatially-varying (SV)BRDF-invariant theory for recovering 3D shape and reflectance from light-field cameras. Our key theoretical insight is a novel analysis of diffuse plus single-lobe SVBRDFs under a light-field setup. We show that, although direct shape recovery is not possible, an equation relating depths and normals can still be derived. Using this equation, we then propose using a polynomial (quadratic) shape prior to resolve the shape ambiguity. Once shape is estimated, we also recover the reflectance. We present extensive synthetic data on the entire MERL BRDF dataset, as well as a number of real examples to validate the theory, where we simultaneously recover shape and BRDFs from a single image taken with a Lytro Illum camera.

We investigate conditional adversarial networks as a general-purpose solution to image-to-image translation problems. These networks not only learn the mapping from input image to output image, but also learn a loss function to train this mapping. This makes it possible to apply the same generic approach to problems that traditionally would require very different loss formulations. We demonstrate that this approach is effective at synthesizing photos from label maps, reconstructing objects from edge maps, and colorizing images, among other tasks. Moreover, since the release of the pi×2pi× software associated with this paper, hundreds of twitter users have posted their own artistic experiments using our system. As a community, we no longer hand-engineer our mapping functions, and this work suggests we can achieve reasonable results without handengineering our loss functions either.

We present a learning framework for abstracting complex shapes by learning to assemble objects using 3D volumetric primitives. In addition to generating simple and geometrically interpretable explanations of 3D objects, our framework also allows us to automatically discover and exploit consistent structure in the data. We demonstrate that using our method allows predicting shape representations which can be leveraged for obtaining a consistent parsing across the instances of a shape collection and constructing an interpretable shape similarity measure. We also examine applications for image-based prediction as well as shape manipulation.

We study the notion of consistency between a 3D shape and a 2D observation and propose a differentiable formulation which allows computing gradients of the 3D shape given an observation from an arbitrary view. We do so by reformulating view consistency using a differentiable ray consistency (DRC) term. We show that this formulation can be incorporated in a learning framework to leverage different types of multi-view observations e.g., foreground masks, depth, color images, semantics etc. as supervision for learning single-view 3D prediction. We present empirical analysis of our technique in a controlled setting. We also show that this approach allows us to improve over existing techniques for single-view reconstruction of objects from the PASCAL VOC dataset.

We propose split-brain autoencoders, a straightforward modification of the traditional autoencoder architecture, for unsupervised representation learning. The method adds a split to the network, resulting in two disjoint sub-networks. Each sub-network is trained to perform a difficult task - predicting one subset of the data channels from another. Together, the sub-networks extract features from the entire input signal. By forcing the network to solve crosschannel prediction tasks, we induce a representation within the network which transfers well to other, unseen tasks. This method achieves state-of-the-art performance on several large-scale transfer learning benchmarks.

Image-to-image translation is a class of vision and graphics problems where the goal is to learn the mapping between an input image and an output image using a training set of aligned image pairs. However, for many tasks, paired training data will not be available. We present an approach for learning to translate an image from a source domain X to a target domain Y in the absence of paired examples. Our goal is to learn a mapping G : X → Y such that the distribution of images from G(X) is indistinguishable from the distribution Y using an adversarial loss. Because this mapping is highly under-constrained, we couple it with an inverse mapping F : Y → X and introduce a cycle consistency loss to push F(G(X)) ≈ X (and vice versa). Qualitative results are presented on several tasks where paired training data does not exist, including collection style transfer, object transfiguration, season transfer, photo enhancement, etc. Quantitative comparisons against several prior methods demonstrate the superiority of our approach.

In many real-world scenarios, rewards extrinsic to the agent are extremely sparse, or absent altogether. In such cases, curiosity can serve as an intrinsic reward signal to enable the agent to explore its environment and learn skills that might be useful later in its life. We formulate curiosity as the error in an agent’s ability to predict the consequence of its own actions in a visual feature space learned by a self-supervised inverse dynamics model. Our formulation scales to high-dimensional continuous state spaces like images, bypasses the difficulties of directly predicting pixels, and, critically, ignores the aspects of the environment that cannot affect the agent. The proposed approach is evaluated in two environments: VizDoom and Super Mario Bros. Three broad settings are investigated: 1) sparse extrinsic reward, where curiosity allows for far fewer interactions with the environment to reach the goal; 2) exploration with no extrinsic reward, where curiosity pushes the agent to explore more efficiently; and 3) generalization to unseen scenarios (e.g. new levels of the same game) where the knowledge gained from earlier experience helps the agent explore new places much faster than starting from scratch.

Many image-to-image translation problems are ambiguous, as a single input image may correspond to multiple possible outputs. In this work, we aim to model a distribution of possible outputs in a conditional generative modeling setting. The ambiguity of the mapping is distilled in a low-dimensional latent vector, which can be randomly sampled at test time. A generator learns to map the given input, combined with this latent code, to the output. We explicitly encourage the connection between output and the latent code to be invertible. This helps prevent a many-to-one mapping from the latent code to the output during training, also known as the problem of mode collapse, and produces more diverse results. We explore several variants of this approach by employing different training objectives, network architectures, and methods of injecting the latent code. Our proposed method encourages bijective consistency between the latent encoding and output modes. We present a systematic comparison of our method and other variants on both perceptual realism and diversity.

Light field cameras have many advantages over traditional cameras, as they allow the user to change various camera settings after capture. However, capturing light fields requires a huge bandwidth to record the data: a modern light field camera can only take three images per second. This prevents current consumer light field cameras from capturing light field videos. Temporal interpolation at such extreme scale (10x, from 3 fps to 30 fps) is infeasible as too much information will be entirely missing between adjacent frames. Instead, we develop a hybrid imaging system, adding another standard video camera to capture the temporal information. Given a 3 fps light field sequence and a standard 30 fps 2D video, our system can then generate a full light field video at 30 fps. We adopt a learning-based approach, which can be decomposed into two steps: spatio-temporal flow estimation and appearance estimation. The flow estimation propagates the angular information from the light field sequence to the 2D video, so we can warp input images to the target view. The appearance estimation then combines these warped images to output the final pixels. The whole process is trained end-to-end using convolutional neural networks. Experimental results demonstrate that our algorithm outperforms current video interpolation methods, enabling consumer light field videography, and making applications such as refocusing and parallax view generation achievable on videos for the first time.

We propose a deep learning approach for user-guided image colorization. The system directly maps a grayscale image, along with sparse, local user "hints" to an output colorization with a Convolutional Neural Network (CNN). Rather than using hand-defined rules, the network propagates user edits by fusing low-level cues along with high-level semantic information, learned from large-scale data. We train on a million images, with simulated user inputs. To guide the user towards efficient input selection, the system recommends likely colors based on the input image and current user inputs. The colorization is performed in a single feed-forward pass, enabling real-time use. Even with randomly simulated user inputs, we show that the proposed system helps novice users quickly create realistic colorizations, and offers large improvements in colorization quality with just a minute of use. In addition, we demonstrate that the framework can incorporate other user "hints" to the desired colorization, showing an application to color histogram transfer.

We present an unsupervised visual feature learning algorithm driven by context-based pixel prediction. By analogy with auto-encoders, we propose Context Encoders - a convolutional neural network trained to generate the contents of an arbitrary image region conditioned on its surroundings. In order to succeed at this task, context encoders need to both understand the content of the entire image, as well as produce a plausible hypothesis for the missing part(s). When training context encoders, we have experimented with both a standard pixel-wise reconstruction loss, as well as a reconstruction plus an adversarial loss. The latter produces much sharper results because it can better handle multiple modes in the output. We found that a context encoder learns a representation that captures not just appearance but also the semantics of visual structures. We quantitatively demonstrate the effectiveness of our learned features for CNN pre-training on classification, detection, and segmentation tasks. Furthermore, context encoders can be used for semantic inpainting tasks, either stand-alone or as initialization for non-parametric methods.

Light-field cameras have recently emerged as a powerful tool for one-shot passive 3D shape capture. However, obtaining the shape of glossy objects like metals, plastics or ceramics remains challenging, since standard Lambertian cues like photo-consistency cannot be easily applied. In this paper, we derive a spatially-varying (SV)BRDF-invariant theory for recovering 3D shape and reflectance from light-field cameras. Our key theoretical insight is a novel analysis of diffuse plus single-lobe SVBRDFs under a light-field setup. We show that, although direct shape recovery is not possible, an equation relating depths and normals can still be derived. Using this equation, we then propose using a polynomial (quadratic) shape prior to resolve the shape ambiguity. Once shape is estimated, we also recover the reflectance. We present extensive synthetic data on the entire MERL BRDF dataset, as well as a number of real examples to validate the theory, where we simultaneously recover shape and BRDFs from a single image taken with a Lytro Illum camera.

Discriminative deep learning approaches have shown impressive results for problems where human-labeled ground truth is plentiful, but what about tasks where labels are difficult or impossible to obtain? This paper tackles one such problem: establishing dense visual correspondence across different object instances. For this task, although we do not know what the ground-truth is, we know it should be consistent across instances of that category. We exploit this consistency as a supervisory signal to train a convolutional neural network to predict cross-instance correspondences between pairs of images depicting objects of the same category. For each pair of training images we find an appropriate 3D CAD model and render two synthetic views to link in with the pair, establishing a correspondence flow 4-cycle. We use ground-truth synthetic-to-synthetic correspondences, provided by the rendering engine, to train a ConvNet to predict synthetic-to-real, real-to-real and real-to-synthetic correspondences that are cycle-consistent with the ground-truth. At test time, no CAD models are required. We demonstrate that our end-to-end trained ConvNet supervised by cycle-consistency outperforms state-of-the-art pairwise matching methods in correspondence-related tasks.

We record, and analyze, and present to the community, KrishnaCam, a large (7.6 million frames, 70 hours) egocentric video stream along with GPS position, acceleration and body orientation data spanning nine months of the life of a computer vision graduate student. We explore and exploit the inherent redundancies in this rich visual data stream to answer simple scene understanding questions such as: How much novel visual information does the student see each day? Given a single egocentric photograph of a scene, can we predict where the student might walk next? We find that given our large video database, simple, nearest-neighbor methods are surprisingly adept baselines for these tasks, even in scenes and scenarios where the camera wearer has never been before. For example, we demonstrate the ability to predict the near-future trajectory of the student in broad set of outdoor situations that includes following sidewalks, stopping to wait for a bus, taking a daily path to work, and the lack of movement while eating food.

Light-field cameras have become widely available in both consumer and industrial applications. However, most previous approaches do not model occlusions explicitly, and therefore fail to capture sharp object boundaries. A common assumption is that for a Lambertian scene, a pixel will exhibit photo-consistency, which means all viewpoints converge to a single point when focused to its depth. However, in the presence of occlusions this assumption fails to hold, making most current approaches unreliable precisely where accurate depth information is most important - at depth discontinuities. In this paper, an occlusion-aware depth estimation algorithm is developed; the method also enables identification of occlusion edges, which may be useful in other applications. It can be shown that although photo-consistency is not preserved for pixels at occlusions, it still holds in approximately half the viewpoints. Moreover, the line separating the two view regions (occluded object versus occluder) has the same orientation as that of the occlusion edge in the spatial domain. By ensuring photo-consistency in only the occluded view region, depth estimation can be improved. Occlusion predictions can also be computed and used for regularization. Experimental results show that our method outperforms current state-of-the-art light-field depth estimation algorithms, especially near occlusion boundaries.

Given a set of poorly aligned images of the same visual concept without any annotations, we propose an algorithm to jointly bring them into pixel-wise correspondence by estimating a FlowWeb representation of the image set. FlowWeb is a fully-connected correspondence flow graph with each node representing an image, and each edge representing the correspondence flow field between a pair of images, i.e. a vector field indicating how each pixel in one image can find a corresponding pixel in the other image. Correspondence flow is related to optical flow but allows for correspondences between visually dissimilar regions if there is evidence they correspond transitively on the graph. Our algorithm starts by initializing all edges of this complete graph with an off-the-shelf, pairwise flow method. We then iteratively update the graph to force it to be more self-consistent. Once the algorithm converges, dense, globally-consistent correspondences can be read off the graph. Our results suggest that FlowWeb improves alignment accuracy over previous pairwise as well as joint alignment methods.

With vast quantities of imagery now available online, researchers have begun to explore whether visual patterns can be discovered automatically. Here we consider the particular domain of architecture, using huge collections of street-level imagery to find visual patterns that correspond to semantic-level architectural elements distinctive to particular time periods. We use this analysis both to date buildings, as well as to discover how functionally-similar architectural elements (e.g. windows, doors, balconies, etc.) have changed over time due to evolving styles. We validate the methods by combining a large dataset of nearly 150,000 Google Street View images from Paris with a cadastre map to infer approximate construction date for each facade. Not only could our analysis be used for dating or geo- localizing buildings based on architectural features, but it also could give architects and historians new tools for confirming known theories or even discovering new ones.

This work explores the use of spatial context as a source of free and plentiful supervisory signal for training a rich visual representation. Given only a large, unlabeled image collection, we extract random pairs of patches from each image and train a convolutional neural net to predict the position of the second patch relative to the first. We argue that doing well on this task requires the model to learn to recognize objects and their parts. We demonstrate that the feature representation learned using this within-image context indeed captures visual similarity across images. For example, this representation allows us to perform unsupervised visual discovery of objects like cats, people, and even birds from the Pascal VOC 2011 detection dataset. Furthermore, we show that the learned ConvNet can be used in the R-CNN framework [19] and provides a significant boost over a randomly-initialized ConvNet, resulting in state-of-the-art performance among algorithms which use only Pascal-provided training set annotations.

Consumer-level and high-end light-field cameras are now widely available. Recent work has demonstrated practical methods for passive depth estimation from light-field images. However, most previous approaches do not explicitly model occlusions, and therefore cannot capture sharp transitions around object boundaries. A common assumption is that a pixel exhibits photo-consistency when focused to its correct depth, i.e., all viewpoints converge to a single (Lambertian) point in the scene. This assumption does not hold in the presence of occlusions, making most current approaches unreliable precisely where accurate depth information is most important - at depth discontinuities. In this paper, we develop a depth estimation algorithm that treats occlusion explicitly, the method also enables identification of occlusion edges, which may be useful in other applications. We show that, although pixels at occlusions do not preserve photo-consistency in general, they are still consistent in approximately half the viewpoints. Moreover, the line separating the two view regions (correct depth vs. occluder) has the same orientation as the occlusion edge has in the spatial domain. By treating these two regions separately, depth estimation can be improved. Occlusion predictions can also be computed and used for regularization. Experimental results show that our method outperforms current state-of-the-art light-field depth estimation algorithms, especially near occlusion boundaries.

We propose a data-driven approach for intrinsic image decomposition, which is the process of inferring the confounding factors of reflectance and shading in an image. We pose this as a two-stage learning problem. First, we train a model to predict relative reflectance ordering between image patches ('brighter', 'darker', 'same') from large-scale human annotations, producing a data-driven reflectance prior. Second, we show how to naturally integrate this learned prior into existing energy minimization frame-works for intrinsic image decomposition. We compare our method to the state-of-the-art approach of Bell et al. [7] on both decomposition and image relighting tasks, demonstrating the benefits of the simple relative reflectance prior, especially for scenes under challenging lighting conditions.

What makes an image appear realistic? In this work, we are answering this question from a data-driven perspective by learning the perception of visual realism directly from large amounts of data. In particular, we train a Convolutional Neural Network (CNN) model that distinguishes natural photographs from automatically generated composite images. The model learns to predict visual realism of a scene in terms of color, lighting and texture compatibility, without any human annotations pertaining to it. Our model outperforms previous works that rely on hand-crafted heuristics, for the task of classifying realistic vs. unrealistic photos. Furthermore, we apply our learned model to compute optimal parameters of a compositing method, to maximize the visual realism score predicted by our CNN model. We demonstrate its advantage against existing methods via a human perception study.

Abstract Recent color transfer methods use local information to learn the transformation from a source to an exemplar image, and then transfer this appearance change to a target image. These solutions achieve very successful results for general mood changes, e.g., changing the appearance of an image from “sunny” to “overcast”. However, such methods have a hard time creating new image content, such as leaves on a bare tree. Texture transfer, on the other hand, can synthesize such content but tends to destroy image structure. We propose the first algorithm that unifies color and texture transfer, outperforming both by leveraging their respective strengths. A key novelty in our approach resides in teasing apart appearance changes that can be modeled simply as changes in color versus those that require new image content to be generated. Our method starts with an analysis phase which evaluates the success of color transfer by comparing the exemplar with the source. This analysis then drives a selective, iterative texture transfer algorithm that simultaneously predicts the success of color transfer on the target and synthesizes new content where needed. We demonstrate our unified algorithm by transferring large temporal changes between photographs, such as change of season – e.g., leaves on bare trees or piles of snow on a street – and flooding.

We propose a data-driven approach to estimate the likelihood that an image segment corresponds to a scene object (its "objectness") by comparing it to a large collection of example object regions. We demonstrate that when the application domain is known, for example, in our case activity of daily living (ADL), we can capture the regularity of the domain specific objects using millions of exemplar object regions. Our approach to estimating the objectness of an image region proceeds in two steps: 1) finding the exemplar regions that are the most similar to the input image segment; 2) calculating the objectness of the image segment by combining segment properties, mutual consistency across the nearest exemplar regions, and the prior probability of each exemplar region. In previous work, parametric objectness models were built from a small number of manually annotated objects regions, instead, our data-driven approach uses 5 million object regions along with their metadata information. Results on multiple data sets demonstrates our data-driven approach compared to the existing model based techniques. We also show the application of our approach in improving the performance of object discovery algorithms.

This paper poses object category detection in images as a type of 2D-to-3D alignment problem, utilizing the large quantities of 3D CAD models that have been made publicly available online. Using the "chair" class as a running example, we propose an exemplar-based 3D category representation, which can explicitly model chairs of different styles as well as the large variation in viewpoint. We develop an approach to establish part-based correspondences between 3D CAD models and real photographs. This is achieved by (i) representing each 3D model using a set of view-dependent mid-level visual elements learned from synthesized views in a discriminative fashion, (ii) carefully calibrating the individual element detectors on a common dataset of negative images, and (iii) matching visual elements to the test image allowing for small mutual deformations but preserving the viewpoint and style constraints. We demonstrate the ability of our system to align 3D models with 2D objects in the challenging PASCAL VOC images, which depict a wide variety of chairs in complex scenes.

This paper addresses the well-established problem of unsupervised object discovery with a novel method inspired by weakly-supervised approaches. In particular, the ability of an object patch to predict the rest of the object (its context) is used as supervisory signal to help discover visually consistent object clusters. The main contributions of this work are: 1) framing unsupervised clustering as a leave-one-out context prediction task; 2) evaluating the quality of context prediction by statistical hypothesis testing between thing and stuff appearance models; and 3) an iterative region prediction and context alignment approach that gradually discovers a visual object cluster together with a segmentation mask and fine-grained correspondences. The proposed method outperforms previous unsupervised as well as weakly-supervised object discovery approaches, and is shown to provide correspondences detailed enough to transfer keypoint annotations.

We describe a method for providing feedback on portrait expressions, and for selecting the most attractive expressions from large video/photo collections. We capture a video of a subject's face while they are engaged in a task designed to elicit a range of positive emotions. We then use crowdsourcing to score the captured expressions for their attractiveness. We use these scores to train a model that can automatically predict attractiveness of different expressions of a given person. We also train a cross-subject model that evaluates portrait attractiveness of novel subjects and show how it can be used to automatically mine attractive photos from personal photo collections. Furthermore, we show how, with a little bit ($5-worth) of extra crowdsourcing, we can substantially improve the cross-subject model by "fine-tuning" it to a new individual using active learning. Finally, we demonstrate a training app that helps people learn how to mimic their best expressions.

Photo-editing software restricts the control of objects in a photograph to the 2D image plane. We present a method that enables users to perform the full range of 3D manipulations, including scaling, rotation, translation, and nonrigid deformations, to an object in a photograph. As 3D manipulations often reveal parts of the object that are hidden in the original photograph, our approach uses publicly available 3D models to guide the completion of the geometry and appearance of the revealed areas of the object. The completion process leverages the structure and symmetry in the stock 3D model to factor out the effects of illumination, and to complete the appearance of the object. We demonstrate our system by producing object manipulations that would be impossible in traditional 2D photo-editing programs, such as turning a car over, making a paper-crane flap its wings, or manipulating airplanes in a historical photograph to change its story.

This paper proposes an interactive framework that allows a user to rapidly explore and visualize a large image collection using the medium ofaverage images. Average images have been gaining popularity as means of artistic expression and data visualization, but the creation of compelling examples is a surprisingly laborious and manual process. Our interactive, real-time system provides a way to summarize large amounts of visual data byweightedaverage(s) of an image collection, with the weights reflecting user-indicated importance. The aim is to capture not just the mean of the distribution, but a set of modes discovered via interactive exploration. We pose this exploration in terms of a user interactively "editing" the average image using various types of strokes, brushes and warps, similar to a normal image editor, with each user interaction providing a new constraint to update the average. New weighted averages can be spawned and edited either individually or jointly. Together, these tools allow the user to simultaneously perform two fundamental operations on visual data: user-guided clustering and user-guided alignment, within the same framework. We show that our system is useful for various computer vision and graphics applications.

We present a method for automatically identifying and validating predictive relationships between the visual appearance of a city and its non-visual attributes (e.g. crime statistics, housing prices, population density etc.). Given a set of street-level images and (location, city-attribute-value) pairs of measurements, we first identify visual elements in the images that are discriminative of the attribute. We then train a predictor by learning a set of weights over these elements using non-linear Support Vector Regression. To perform these operations efficiently, we implement a scalable distributed processing framework that speeds up the main computational bottleneck (extracting visual elements) by an order of magnitude. This speedup allows us to investigate a variety of city attributes across 6 different American cities. We find that indeed there is a predictive relationship between visual elements and a number of city attributes including violent crime rates, theft rates, housing prices, population density, tree presence, graffiti presence, and the perception of danger. We also test human performance for predicting theft based on street-level images and show that our predictor outperforms this baseline with 33% higher accuracy on average. Finally, we present three prototype applications that use our system to (1) define the visual boundary of city neighborhoods, (2) generate walking directions that avoid or seek out exposure to city attributes, and (3) validate user-specified visual elements for prediction.

We present a weakly-supervised visual data mining approach that discovers connections between recurring mid-level visual elements in historic (temporal) and geographic (spatial) image collections, and attempts to capture the underlying visual style. In contrast to existing discovery methods that mine for patterns that remain visually consistent throughout the dataset, our goal is to discover visual elements whose appearance changes due to change in time or location; i.e., exhibit consistent stylistic variations across the label space (date or geo-location). To discover these elements, we first identify groups of patches that are style-sensitive. We then incrementally build correspondences to find the same element across the entire dataset. Finally, we train style-aware regressors that model each element's range of stylistic differences. We apply our approach to date and geo-location prediction and show substantial improvement over several baselines that do not model visual style. We also demonstrate the method's effectiveness on the related task of fine-grained classification.

Recent work on mid-level visual representations aims to capture information at the level of complexity higher than typical visual words", but lower than full-blown semantic objects. Several approaches have been proposed to discover mid-level visual elements, that are both 1) representative, i.e. frequently occurring within a visual dataset, and 2) visually discriminative. However, the current approaches are rather ad hoc and difficult to analyze and evaluate. In this work, we pose visual element discovery as discriminative mode seeking, drawing connections to the the well-known and well-studied mean-shift algorithm. Given a weakly-labeled image collection, our method discovers visually-coherent patch clusters that are maximally discriminative with respect to the labels. One advantage of our formulation is that it requires only a single pass through the data. We also propose the Purity-Coverage plot as a principled way of experimentally analyzing and evaluating different visual discovery approaches, and compare our method against prior work on the Paris Street View dataset. We also evaluate our method on the task of scene classification, demonstrating state-of-the-art performance on the MIT Scene-67 dataset."

Traditionally, computer graphics has been concerned with producing imagery that is as physically accurate as possible. But accurate physical simulation of geometry, lighting, and material properties of a visual scene can be cumbersome and time consuming. At the same time, human vision is far from accurate, which offers an enormous opportunity to create imagery at a reduced computational cost as well as with less reliance on human modelers. As a result, a recent trend is toward accepting perceptual plausibility instead of physical accuracy as a guiding principle in the design of modeling and rendering systems. This requires us to understand visual realism, which involves both learning statistical regularities of the world, for instance, by employing huge amounts of data, as well as human's visual perception of it. This paper addresses issues related to understanding realism, presents several applications, and discusses what this interesting approach may lead to in the future.

Most contemporary object detection approaches assume each object instance in the training data to be uniquely represented by a single bounding box. In this paper, we go beyond this conventional view by allowing an object instance to be described by multiple bounding boxes. The new bounding box annotations are determined based on the alignment of an object instance with the other training instances in the dataset. Our proposal enables the training data to be reused multiple times for training richer multi-component category models. We operationalize this idea by two complementary operations: bounding box shrinking, which finds subregions of an object instance that could be shared; and bounding box enlarging, which enlarges object instances to include local contextual cues. We empirically validate our approach on the PASCAL VOC detection dataset.

Given a large repository of geotagged imagery, we seek to automatically find visual elements, e. g. windows, balconies, and street signs, that are most distinctive for a certain geo-spatial area, for example the city of Paris. This is a tremendously difficult task as the visual features distinguishing architectural elements of different places can be very subtle. In addition, we face a hard search problem: given all possible patches in all images, which of them are both frequently occurring and geographically informative? To address these issues, we propose to use a discriminative clustering approach able to take into account the weak geographic supervision. We show that geographically representative image elements can be discovered automatically from Google Street View imagery in a discriminative manner. We demonstrate that these elements are visually interpretable and perceptually geo-informative. The discovered visual elements can also support a variety of computational geography tasks, such as mapping architectural correspondences and influences within and across cities, finding representative elements at different geo-spatial scales, and geographically-informed image retrieval.

We present a human-centric paradigm for scene understanding. Our approach goes beyond estimating 3D scene geometry and predicts the "workspace" of a human which is represented by a data-driven vocabulary of human interactions. Our method builds upon the recent work in indoor scene understanding and the availability of motion capture data to create a joint space of human poses and scene geometry by modeling the physical interactions between the two. This joint space can then be used to predict potential human poses and joint locations from a single image. In a way, this work revisits the principle of Gibsonian affordances, reinterpreting it for the modern, data-driven era.

Datasets are an integral part of contemporary object recognition research. They have been the chief reason for the considerable progress in the field, not just as source of large amounts of training data, but also as means of measuring and comparing performance of competing algorithms. At the same time, datasets have often been blamed for narrowing the focus of object recognition research, reducing it to a single benchmark performance number. Indeed, some datasets, that started out as data capture efforts aimed at representing the visual world, have become closed worlds unto themselves (e.g. the Corel world, the Caltech-101 world, the PASCAL VOC world). With the focus on beating the latest benchmark numbers on the latest dataset, have we perhaps lost sight of the original purpose? The goal of this paper is to take stock of the current state of recognition datasets. We present a comparison study using a set of popular datasets, evaluated based on a number of criteria including: relative data bias, cross-dataset generalization, effects of closed-world assumption, and sample value. The experimental results, some rather surprising, suggest directions that can improve dataset collection as well as algorithm evaluation protocols. But more broadly, the hope is to stimulate discussion in the community regarding this very important, but largely neglected issue.

This paper proposes a conceptually simple but surprisingly powerful method which combines the effectiveness of a discriminative object detector with the explicit correspondence offered by a nearest-neighbor approach. The method is based on training a separate linear SVM classifier for every exemplar in the training set. Each of these Exemplar-SVMs is thus defined by a single positive instance and millions of negatives. While each detector is quite specific to its exemplar, we empirically observe that an ensemble of such Exemplar-SVMs offers surprisingly good generalization. Our performance on the PASCAL VOC detection task is on par with the much more complex latent part-based model of Felzenszwalb et al., at only a modest computational cost increase. But the central benefit of our approach is that it creates an explicit association between each detection and a single training exemplar. Because most detections show good alignment to their associated exemplar, it is possible to transfer any available exemplar meta-data (segmentation, geometric structure, 3D model, etc.) directly onto the detections, which can then be used as part of overall scene understanding.

The goal of this work is to find visually similar images even if they appear quite different at the raw pixel level. This task is particularly important for matching images across visual domains, such as photos taken over different seasons or lighting conditions, paintings, hand-drawn sketches, etc. We propose a surprisingly simple method that estimates the relative importance of different features in a query image based on the notion of "data-driven uniqueness". We employ standard tools from discriminative object detection in a novel way, yielding a generic approach that does not depend on a particular image representation or a specific visual domain. Our approach shows good performance on a number of difficult cross-domain visual tasks e.g., matching paintings or sketches to real photographs. The method also allows us to demonstrate novel applications such as Internet re-photography , and painting2gps. While at present the technique is too computationally intensive to be practical for interactive image retrieval, we hope that some of the ideas will eventually become applicable to that domain as well.

This paper presents an empirical evaluation of the role of context in a contemporary, challenging object detection task - the PASCAL VOC 2008. Previous experiments with context have mostly been done on home-grown datasets, often with non-standard baselines, making it difficult to isolate the contribution of contextual information. In this work, we present our analysis on a standard dataset, using top-performing local appearance detectors as baseline. We evaluate several different sources of context and ways to utilize it. While we employ many contextual cues that have been used before, we also propose a few novel ones including the use of geographic context and a new approach for using object spatial support.

This paper presents a method for estimating geographic location for sequences of time-stamped photographs. A prior distribution over travel describes the likelihood of traveling from one location to another during a given time interval. This distribution is based on a training database of 6 million photographs from Flickr.com. An image likelihood for each location is defined by matching a test photograph against the training database. Inferring location for images in a test sequence is then performed using the Forward-Backward algorithm, and the model can be adapted to individual users as well. Using temporal constraints allows our method to geolocate images without recognizable landmarks, and images with no geographic cues whatsoever. This method achieves a substantial performance improvement over the best-available baseline, and geolocates some users' images with near-perfect accuracy.

Given a single outdoor image, we present a method for estimating the likely illumination conditions of the scene. In particular, we compute the probability distribution over the sun position and visibility. The method relies on a combination of weak cues that can be extracted from different portions of the image: the sky, the vertical surfaces, and the ground. While no single cue can reliably estimate illumination by itself, each one can reinforce the others to yield a more robust estimate. This is combined with a data-driven prior computed over a dataset of 6 million Internet photos. We present quantitative results on a webcam dataset with annotated sun positions, as well as qualitative results on consumer-grade photographs downloaded from Internet. Based on the estimated illumination, we show how to realistically insert synthetic 3-D objects into the scene.

The use of context is critical for scene understanding in computer vision, where the recognition of an object is driven by both local appearance and the objects relationship to other elements of the scene (context). Most current approaches rely on modeling the relationships between object categories as a source of context. In this paper we seek to move beyond categories to provide a richer appearance-based model of context. We present an exemplar-based model of objects and their relationships, the Visual Memex, that encodes both local appearance and 2D spatial context between object instances. We evaluate our model on Torralbas proposed Context Challenge against a baseline category-based system. Our experiments suggest that moving beyond categories for context modeling appears to be quite beneficial, and may be the critical missing ingredient in scene understanding systems.

In this paper, we investigate how similar images sharing the same global description can help with unsupervised scene segmentation in an image. In contrast to recent work in semantic alignment of scenes, we allow an input image to be explained by partial matches of similar scenes. This allows for a better explanation of the input scenes. We perform MRF-based segmentation that optimizes over matches, while respecting boundary information. The recovered segments are then used to re-query a large database of images to retrieve better matches for the target region. We show improved performance in detecting occluding boundaries over previous methods on data gathered from the LabelMe database.

Webcams placed all over the world observe and record the visual appearance of a variety of outdoor scenes over long periods of time. The recorded time-lapse image sequences cover a wide range of illumination and weather conditions -- a vast untapped resource for creating visual realism. In this work, we propose to use a large repository of webcams as a "clip art" library from which users may transfer scene appearance (objects, scene backdrops, outdoor illumination) into their own time-lapse sequences or even single photographs. The goal is to combine the recent ideas from data-driven appearance transfer techniques with a general and theoretically-grounded physically-based illumination model. To accomplish this, the paper presents three main research contributions: 1) a new, high-quality outdoor webcam database that has been calibrated radiometrically and geometrically; 2) a novel approach for matching illuminations across different scenes based on the estimation of the properties of natural illuminants (sun, sky, weather and clouds), the camera geometry, and illumination-dependent scene features; 3) a new algorithm for generating physically plausible high dynamic range environment maps for each frame in a webcam sequence.

Estimating geographic information from an image is an excellent, difficult high-level computer vision problem whose time has come. The emergence of vast amounts of geographically-calibrated image data is a great reason for computer vision to start looking globally - on the scale of the entire planet! In this paper, we propose a simple algorithm for estimating a distribution over geographic locations from a single image using a purely data-driven scene matching approach. For this task, we leverage a dataset of over 6 million GPS-tagged images from the Internet. We represent the estimated image location as a probability distribution over the Earthpsilas surface. We quantitatively evaluate our approach in several geolocation tasks and demonstrate encouraging performance (up to 30 times better than chance). We show that geolocation estimates can provide the basis for numerous other image understanding tasks such as population density estimation, land cover estimation or urban/rural classification.

Image understanding involves analyzing many different aspects of the scene. In this paper, we are concerned with how these tasks can be combined in a way that improves the performance of each of them. Inspired by Barrow and Tenenbaum, we present a flexible framework for interfacing scene analysis processes using intrinsic images. Each intrinsic image is a registered map describing one characteristic of the scene. We apply this framework to develop an integrated 3D scene understanding system with estimates of surface orientations, occlusion boundaries, objects, camera viewpoint, and relative depth. Our experiments on a set of 300 outdoor images demonstrate that these tasks reinforce each other, and we illustrate a coherent scene understanding with automatically reconstructed 3D models.

We pose the recognition problem as data association. In this setting, a novel object is explained solely in terms of a small set of exemplar objects to which it is visually similar. Inspired by the work of Frome et al., we learn separate distance functions for each exemplar; however, our distances are interpretable on an absolute scale and can be thresholded to detect the presence of an object. Our exemplars are represented as image regions and the learned distances capture the relative importance of shape, color, texture, and position features for that region. We use the distance functions to detect and segment objects in novel images by associating the bottom-up segments obtained from multiple image segmentations with the exemplar regions. We evaluate the detection and segmentation performance of our algorithm on real-world outdoor scenes from the LabelMe (B. Russel, et al., 2007) dataset and also show some promising qualitative image parsing results.

Objects in the world can be arranged into a hierarchy based on their semantic meaning (e.g. organism - animal - feline - cat). What about defining a hierarchy based on the visual appearance of objects? This paper investigates ways to automatically discover a hierarchical structure for the visual world from a collection of unlabeled images. Previous approaches for unsupervised object and scene discovery focused on partitioning the visual data into a set of non-overlapping classes of equal granularity. In this work, we propose to group visual objects using a multi-layer hierarchy tree that is based on common visual elements. This is achieved by adapting to the visual domain the generative hierarchical latent Dirichlet allocation (hLDA) model previously used for unsupervised discovery of topic hierarchies in text. Images are modeled using quantized local image regions as analogues to words in text. Employing the multiple segmentation framework of Russell et al. [22], we show that meaningful object hierarchies, together with object segmentations, can be automatically learned from unlabeled and unsegmented image collections without supervision. We demonstrate improved object classification and localization performance using hLDA over the previous non-hierarchical method on the MSRC dataset [33].

Abstract Many categories of objects, such as human faces, can be naturally viewed as a composition of several different layers. For example, a bearded face with glasses can be decomposed into three layers: a layer for glasses, a layer for the beard and a layer for other permanent facial features. While modeling such a face with a linear subspace model could be very difficult, layer separation allows for easy modeling and modification of some certain structures while leaving others unchanged. In this paper, we present a method for automatic layer extraction and its applications to face synthesis and editing. Layers are automatically extracted by utilizing the differences between subspaces and modeled separately. We show that our method can be used for tasks such beard removal (virtual shaving), beard synthesis, and beard transfer, among others.

Sliding window scanning is the dominant paradigm in object recognition research today. But while much success has been reported in detecting several rectangular-shaped object classes (i.e. faces, cars, pedestrians), results have been much less impressive for more general types of objects. Several researchers have advocated the use of image segmentation as a way to get a better spatial support for objects. In this paper, our aim is to address this issue by studying the following two questions: 1) how important is good spatial support for recognition? 2) can segmentation provide better spatial support for objects? To answer the first, we compare recognition performance using ground-truth segmentation vs. bounding boxes. To answer the second, we use the multiple segmentation approach to evaluate how close can real segments approach the ground-truth for real objects, and at what cost. Our results demonstrate the importance of finding the right spatial support for objects, and the feasibility of doing so without excessive computational burden.

Occlusion reasoning, necessary for tasks such as navigation and object search, is an important aspect of everyday life and a fundamental problem in computer vision. We believe that the amazing ability of humans to reason about occlusions from one image is based on an intrinsically 3D interpretation. In this paper, our goal is to recover the occlusion boundaries and depth ordering of free-standing structures in the scene. Our approach is to learn to identify and label occlusion boundaries using the traditional edge and region cues together with 3D surface and depth cues. Since some of these cues require good spatial support (i.e., a segmentation), we gradually create larger regions and use them to improve inference over the boundaries. Our experiments demonstrate the power of a scene-based approach to occlusion reasoning.

Why does placing an object from one photograph into another often make the colors of that object suddenly look wrong? One possibility is that humans prefer distributions of colors that are often found in nature; that is, we find pleasing these color combinations that we see often. Another possibility is that humans simply prefer colors to be consistent within an image, regardless of what they are. In this paper, we explore some of these issues by studying the color statistics of a large dataset of natural images, and by looking at differences in color distribution in realistic and unrealistic images. We apply our findings to two problems: 1) classifying composite images into realistic vs. non- realistic, and 2) recoloring image regions for realistic compositing.

What can you do with a million images? In this paper we present a new image completion algorithm powered by a huge database of photographs gathered from the Web. The algorithm patches up holes in images by finding similar image regions in the database that are not only seamless but also semantically valid. Our chief insight is that while the space of images is effectively infinite, the space of semantically differentiable scenes is actually not that large. For many image completion tasks we are able to find similar scenes which contain image fragments that will convincingly complete the image. Our algorithm is entirely data-driven, requiring no annotations or labelling by the user. Unlike existing image completion methods, our algorithm can generate a diverse set of results for each input image and we allow users to select among them. We demonstrate the superiority of our algorithm over existing image completion approaches.

We present a system for inserting new objects into existing photographs by querying a vast image-based object library, pre-computed using a publicly available Internet object database. The central goal is to shield the user from all of the arduous tasks typically involved in image compositing. The user is only asked to do two simple things: 1) pick a 3D location in the scene to place a new object; 2) select an object to insert using a hierarchical menu. We pose the problem of object insertion as a data-driven, 3D-based, context-sensitive object retrieval task. Instead of trying to manipulate the object to change its orientation, color distribution, etc. to fit the new image, we simply retrieve an object of a specified class that has all the required properties (camera pose, lighting, resolution, etc) from our large object library. We present new automatic algorithms for improving object segmentation and blending, estimating true 3D object size and orientation, and estimating scene lighting conditions. We also present an intuitive user interface that makes object insertion fast and simple even for the artistically challenged.

Image understanding requires not only individually estimating elements of the visual world but also capturing the interplay among them. In this paper, we provide a framework for placing local object detection in the context of the overall 3D scene by modeling the interdependence of objects, surface orientations, and camera viewpoint. Most object detection methods consider all scales and locations in the image as equally likely. We show that with probabilistic estimates of 3D geometry, both in terms of surfaces and world coordinates, we can put objects into perspective and model the scale and location variance in the image. Our approach reflects the cyclical nature of the problem by allowing probabilistic object hypotheses to refine geometry and vice-versa. Our framework allows painless substitution of almost any object detector and is easily extended to include other aspects of image understanding. Our results confirm the benefits of our integrated approach.

Given a large dataset of images, we seek to automatically determine the visually similar object and scene classes together with their image segmentation. To achieve this we combine two ideas: (i) that a set of segmented objects can be partitioned into visual object classes using topic discovery models from statistical text analysis; and (ii) that visual object classes can be used to assess the accuracy of a segmentation. To tie these ideas together we compute multiple segmentations of each image and then: (i) learn the object classes; and (ii) choose the correct segmentations. We demonstrate that such an algorithm succeeds in automatically discovering many familiar objects in a variety of image datasets, including those from Caltech, MSRC and LabelMe.

Mobile robots need maps or other forms of geometric information about the environment to navigate. The mobility sensors (LADAR, stereo, etc.) on these robotic vehicles can however populate these maps only up to a distance of a few tens of meters. A navigation system has no knowledge about the world beyond this sensing horizon. As a result, path planners that rely only on this knowledge are unable to anticipate obstacles sufficiently early and have no choice but to resort to an inefficient local obstacle avoidance behavior. However, recent developments in the computer vision community allows us to collect geometric information about the environment far beyond this sensing horizon. The coarse 3D geometric estimation that can be recovered is derived from an appearance-based model. That uses a multiple-hypothesis framework to robustly estimate scene structure from a single image and estimating confidences for each geometric label. This 3D geometric estimation is used with a previously presented navigation strategy that reasons about sensor constraints and plans for measurements while navigating towards the goal. The validity of the sensing method and navigation strategy is supported by results from simulations as well as field experiments with a real robotic platform. These results also show that significant reduction in path length can be achieved by using this framework

Many computer vision algorithms limit their performance by ignoring the underlying 3D geometric structure in the image. We show that we can estimate the coarse geometric properties of a scene by learning appearance-based models of geometric classes, even in cluttered natural scenes. Geometric classes describe the 3D orientation of an image region with respect to the camera. We provide a multiple-hypothesis framework for robustly estimating scene structure from a single image and obtaining confidences for each geometric label. These confidences can then be used to improve the performance of many other applications. We provide a thorough quantitative evaluation of our algorithm on a set of outdoor images and demonstrate its usefulness in two applications: object detection and automatic single-view reconstruction.

We seek to discover the object categories depicted in a set of unlabelled images. We achieve this using a model developed in the statistical text literature: probabilistic latent semantic analysis (pLSA). In text analysis, this is used to discover topics in a corpus using the bag-of-words document representation. Here we treat object categories as topics, so that an image containing instances of several categories is modeled as a mixture of topics. The model is applied to images by using a visual analogue of a word, formed by vector quantizing SIFT-like region descriptors. The topic discovery approach successfully translates to the visual domain: for a small set of objects, we show that both the object categories and their approximate spatial layout are found without supervision. Performance of this unsupervised method is compared to the supervised approach of Fergus et al. (2003) on a set of unseen images containing only one object per image. We also extend the bag-of-words vocabulary to include 'doublets' which encode spatially local co-occurring regions. It is demonstrated that this extended vocabulary gives a cleaner image segmentation. Finally, the classification and segmentation methods are applied to a set of images containing multiple objects per image. These results demonstrate that we can successfully build object class models from an unsupervised analysis of images.

This paper presents a fully automatic method for creating a 3D model from a single photograph. The model is made up of several texture-mapped planar billboards and has the complexity of a typical children's pop-up book illustration. Our main insight is that instead of attempting to recover precise geometry, we statistically modelgeometric classesdefined by their orientations in the scene. Our algorithm labels regions of the input image into coarse categories: "ground", "sky", and "vertical". These labels are then used to "cut and fold" the image into a pop-up model using a set of simple assumptions. Because of the inherent ambiguity of the problem and the statistical nature of the approach, the algorithm is not expected to work on every image. However. it performs surprisingly well for a wide range of scenes taken from a typical person's photo album.

In this paper, we investigate whether it is possible to develop a measure that quantifies the naturalness of human motion (as defined by a large database). Such a measure might prove useful in verifying that a motion editing operation had not destroyed the naturalness of a motion capture clip or that a synthetic motion transition was within the space of those seen in natural human motion. We explore the performance of mixture of Gaussians (MoG), hidden Markov models (HMM), and switching linear dynamic systems (SLDS) on this problem. We use each of these statistical models alone and as part of an ensemble of smaller statistical models. We also implement a Naive Bayes (NB) model for a baseline comparison. We test these techniques on motion capture data held out from a database, keyframed motions, edited motions, motions with noise added, and synthetic motion transitions. We present the results as receiver operating characteristic (ROC) curves and compare the results to the judgments made by subjects in a user study.

We consider the problem of recovering an underwater image distorted by surface waves. A large amount of video data of the distorted image is acquired. The problem is posed in terms of finding an undistorted im- age patch at each spatial location. This challenging reconstruction task can be formulated as a manifold learning problem, such that the center of the manifold is the image of the undistorted patch. To compute the center, we present a new technique to estimate global distances on the manifold. Our technique achieves robustness through convex flow com- putations and solves the "leakage" problem inherent in recent manifold embedding techniques. 1 Introduction Consider the following problem. A pool of water is observed by a stationary video camera mounted above the pool and looking straight down. There are waves on the surface of the water and all the camera sees is a series of distorted images of the bottom of the pool, e.g. Figure 1. The aim is to use these images to recover the undistorted image of the pool floor as if the water was perfectly still. Besides obvious applications in ocean optics and underwater imaging [1], variants of this problem also arise in several other fields, including astronomy (overcoming atmospheric distortions) and structure-from-motion (learning the appearance of a deforming object). Most approaches to solve this problem try to model the distortions explicitly. In order to do this, it is critical not only to have a good parametric model of the distortion process, but also to be able to reliably extract features from the data to fit the parameters. As such, this approach is only feasible in well understood, highly controlled domains. On the opposite side of the spectrum is a very simple method used in underwater imaging: simply, average the data temporally. Although this method performs surprisingly well in many situations, it fails when the structure of the target image is too fine with respect to the amplitude of the wave (Figure 2). In this paper we propose to look at this difficult problem from a more statistical angle. We will exploit a very simple observation: if we watch a particular spot on the image plane, most of the time the picture projected there will be distorted. But once in a while, when the water just happens to be locally flat at that point, we will be looking straight down and seeing exactly the right spot on the ground. If we can recognize when this happens Authors in alphabetical order. Figure 1: Fifteen consecutive frames from the video. The experimental setup involved: a transparent bucket of water, the cover of a vision textbook "Computer Vision/A Modern Approach". Figure 2: Ground truth image and reconstruction results using mean and median and snap the right picture at each spatial location, then recovering the desired ground truth image would be simply a matter of stitching these correct observations together. In other words, the question that we will be exploring in this paper is not where to look, but when! 2 Problem setup Let us first examine the physical setup of our problem. There is a "ground truth" image G on the bottom of the pool. Overhead, a stationary camera pointing downwards is recording a video stream V . In the absence of any distortion V (x, y, t) = G(x, y) at any time t. However, the water surface refracts in accordance with Snell's Law. Let us consider what the camera is seeing at a particular point x on the CCD array, as shown in Figure 3(c) (assume 1D for simplicity). If the normal to the water surface directly underneath x is pointing straight up, there is no refraction and V (x) = G(x). However, if the normal is tilted by angle 1, light will bend by the amount 2 = 1 - sin-1 ( 1 sin 1.33 1 ), so the camera point V (x) will see the light projected from G(x + dx) on the ground plane. It is easy to see that the relationship between the tilt of the normal to the surface 1 and the displacement dx is approximately linear (dx 0.251h using small angle approximation, where h is the height of the water). This means that, in 2D, what the camera will be seeing over time at point V (x, y, t) are points on the ground plane sampled from a disk centered at G(x, y) and with radius related to the height of the water and the overall roughness of the water surface. A similar relationship holds in the inverse direction as well: a point G(x, y) will be imaged on a disk centered around V (x, y). What about the distribution of these sample points? According to Cox-Munk Law [2], the surface normals of rough water are distributed approximately as a Gaussian centered around the vertical, assuming a large surface area and stationary waves. Our own experiments, conducted by hand-tracking (Figure 3b), confirm that the distribution, though not exactly Gaussian, is definitely unimodal and smooth. Up to now, we only concerned ourselves with infinitesimally small points on the image or the ground plane. However, in practice, we must have something that we can compute with. Therefore, we will make an assumption that the surface of the water can be locally approximated by a planar patch. This means that everything that was true for points is now true for local image patches (up to a small affine distortion). 3 Tracking via embedding From the description outlined above, one possible solution emerges. If the distribution of a particular ground point on the image plane is unimodal, then one could track feature points in the video sequence over time. Computing their mean positions over the entire video will give an estimate of their true positions on the ground plane. Unfortunately, tracking over long periods of time is difficult even under favorable conditions, whereas our data is so fast (undersampled) and noisy that reliable tracking is out of the question (Figure 3(c)). However, since we have a lot of data, we can substitute smoothness in time with smoothness in similarity for a given patch we are more likely to find a patch similar to it somewhere in time, and will have a better chance to track the transition between them. An alternative to tracking the patches directly (which amounts to holding the ground patch G(x, y) fixed and centering the image patches V (x + dxt, y + dyt) on top of it in each frame), is to fix the image patch V (x, y) in space and observe the patches from G(x + dxt, y + dyt) appearing in this window. We know that this set of patches comes from a disk on the ground plane centered around patch G(x, y) our goal. If the disk was small enough compared to the size of the patch, we could just cluster the patches together, e.g. by using translational EM [3]. Unfortunately, the disk can be rather large, containing patches with no overlap at all, thus making only the local similarity comparisons possible. However, notice that our set of patches lies on a low-dimensional manifold; in fact we know precisely which manifold it's the disk on the ground plane centered at G(x, y)! So, if we could use the local patch similarities to find an embedding of the patches in V (x, y, t) on this manifold, the center of the embedding will hold our desired patch G(x, y). The problem of embedding the patches based on local similarity is related to the recent work in manifold learning [4, 5]. Basic ingredients of the embedding algorithms are: defin- ing a distance measure between points, and finding an energy function that optimally places them in the embedding space. The distance can be defined as all-pairs distance matrix, or as distance from a particular reference node. In both cases, we want the distance function to satisfy some constraints to model the underlying physical problem. The local similarity measure for our problem turned out to be particularly unreliable, so none of the previous manifold learning techniques were adequate for our purposes. In the following section we will describe our own, robust method for computing a global distance function and finding the right embedding and eventually the center of it. 1 N Surface h 2 G(x) G(x + dx) (a) (b) (c) Figure 3: (a) Snell's Law (b)-(c) Tracking points of the bottom of the pool: (b) the tracked position forms a distribution close to a Gaussian, (c): a vertical line of the image shown at different time instances (horizontal axis). The discontinuity caused by rapid changes makes the tracking infeasible. 4 What is the right distance function? Let I = {I1, . . . , In} be the set of patches, where It = V (x, y, t) and x = [xmin, xmax], y = [ymin, ymax] are the patch pixel coordinates. Our goal is to find a center patch to represent the set I. To achieve this goal, we need a distance function d : I I IR such that d(Ii, Ij) < d(Ii, Ik) implies that Ij is more similar to Ii than Ik. Once we have such a measure, the center can be found by computing: I = arg min d(Ii, Ij) (1) IiI Ij I Unfortunately, the measurable distance functions, such as Normalized Cross Correlation (N CC) are only local. A common approach is to design a global distance function using the measurable local distances and transitivity [6, 4]. This is equivalent to designing a global distance function of the form: d d(I local(Ii, Ij ), if dlocal(Ii, Ij) i, Ij ) = (2) dtransitive(Ii, Ij), otherwise. where dlocal is a local distance function, is a user-specified threshold and dtransitive is a global, transitive distance function which utilizes dlocal. The underlying assumption here is that the members of I lie on a constraint space (or manifold) S. Hence, a local similarity function such as N CC can be used to measure local distances on the manifold. An important research question in machine learning is to extend the local measurements into global ones, i.e. to design dtransitive above. One method for designing such a transitive distance function is to build a graph G = (V, E) whose vertices correspond to the members of I. The local distance measure is used to place edges which connect only very similar members of I. Afterwards, the length of pairwise shortest paths are used to estimate the true distances on the manifold S. For example, this method forms the basis of the well-known Isomap method [4]. Unfortunately, estimating the distance dtransitive(, ) using shortest path computations is not robust to errors in the local distances which are very common. Consider a patch that contains the letter A and another one that contains the letter B. Since they are different letters, we expect that these patches would be quite distant on the manifold S. However, among the A patches there will inevitably be a very blurry A that would look quite similar to a very blurry B producing an erroneous local distance measurement. When the transitive global distances are computed using shortest paths, a single erroneous edge will single- handedly cause all the A patches to be much closer to all the B patches, short-circuiting the graph and completely distorting all the distances. Such errors lead to the leakage problem in estimating the global distances of patches. This problem is illustrated in Figure 4. In this example, our underlying manifold S is a triangle. Suppose our local distance function erroneously estimates an edge between the corners of the triangle as shown in the figure. After the erroneous edge is inserted, the shortest paths from the top of the triangle leak through this edge. Therefore, the shortest path distances will fail to reflect the true distance on the manifold. 5 Solving the leakage problem Recall that our goal is to find the center of our data set as defined in Equation 1. Note that, in order to compute the center we do not need all pairwise distances. All we need is the quantity dI (Ii) = d(I I i, Ij ) for all Ii. j I The leakage problem occurs when we compute the values dI (Ii) using the shortest path metric. In this case, even a single erroneous edge may reduce the shortest paths from many different patches to Ii changing the value of dI(Ii) drastically. Intuitively, in order to prevent the leakage problem we must prevent edges from getting involved in many shortest path computations to the same node (i.e. leaking edges). We can formalize this notion by casting the computation as a network flow problem. Let G = (V, E) be our graph representation such that for each patch Ii I, there is a vertex vi V . The edge set E is built as follows: there is an edge (vi, vj) if dlocal(Ii, Ij) is less than a threshold. The weight of the edge (vi, vj) is equal to dlocal(Ii, Ij). To compute the value dI (Ii), we build a flow network whose vertex set is also V . All vertices in V - {vi} are sources, pushing unit flow into the network. The vertex vi is a sink with infinite capacity. The arcs of the flow network are chosen using the edge set E. For each edge (vj, vk) E we add the arcs vj vk and vk vj. Both arcs have infinite capacity and the cost of pushing one unit of flow on either arc is equal to the weight of (vj, vk), as shown in Figure 4 left (top and bottom). It can easily be seen that the minimum cost flow in this network is equal to dI (Ii). Let us call this network which is used to compute dI (Ii) as N W (Ii). The crucial factor in designing such a flow network is choosing the right cost and capacity. Computing the minimum cost flow on N W (Ii) not only gives us dI(Ii) but also allows us to compute how many times an edge is involved in the distance computation: the amount of flow through an edge is exactly the number of times that edge is used for the shortest path computations. This is illustrated in Figure 4 (box A) where d1 units of cost is charged for each unit of flow through the edge (u, w). Therefore, if we prevent too much flow going through an edge, we can prevent the leakage problem. d3/ d1/ d Error u 2/c2 w u w d1 v A: Shortest Path B: Convex Flow c d1/c1 1 c1 + c2 u C: Shortest Path with Capacity Error d/ d1/c1 v u w v c1 w Figure 4: The leakage problem. Left: Equivalence of shortest path leakage and uncapacitated flow leakage problem. Bottom-middle: After the erroneous edge is inserted, the shortest paths from the top of the triangle to vertex v go through this edge. Boxes A-C:Alternatives for charging a unit of flow between nodes u and w. The horizontal axis of the plots is the amount of flow and the vertical axis is the cost. Box A: Linear flow. The cost of a unit of flow is d1 Box B: Convex flow. Multiple edges are introduced between two nodes, with fixed capacity, and convexly increasing costs. The cost of a unit of flow increases from d1 to d2 and then to d3 as the amount of flow from u to w increases. Box C: Linear flow with capacity. The cost is d1 until a capacity of c1 is achieved and becomes infinite afterwards. One might think that the leakage problem can simply be avoided by imposing capacity constraints on the arcs of the flow network (Figure 4, box C). Unfortunately, this is not very easy. Observe that in the minimum cost flow solution of the network N W (Ii), the amount of flow on the arcs will increase as the arcs get closer to Ii. Therefore, when we are setting up the network N W (Ii), we must adaptively increase the capacities of arcs "closer" to the sink vi otherwise, there will be no feasible solution. As the structure of the graph G gets complicated, specifying this notion of closeness becomes a subtle issue. Further, the structure of the underlying space S could be such that some arcs in G must indeed carry a lot of flow. Therefore imposing capacities on the arcs requires understanding the underlying structure of the graph G as well as the space S which is in fact the problem we are trying to solve! Our proposed solution to the leakage problem uses the notion of a convex flow. We do not impose a capacity on the arcs. Instead, we impose a convex cost function on the arcs such that the cost of pushing unit flow on arc a increases as the total amount of flow through a increases. See Figure 4, box B. This can be achieved by transforming the network N W (Ii) to a new network N W (Ii). The transformation is achieved by applying the following operation on each arc in N W (Ii): Let a be an arc from u to w in N W (Ii). In N W (Ii), we replace a by k arcs a1, . . . , ak. The costs of these arcs are chosen to be uniformly increasing so that cost(a1) < cost(a2) < . . . < cost(ak). The capacity of arc ak is infinite. The weights and capacities of the other arcs are chosen to reflect the steepness of the desired convexity (Figure 4, box B). The network shown in the figure yields the following function for the cost of pushing x units of flow through the arc: d1x, if 0 x c1 cost(x) = d1c1 + d2(x - c1), if c1 x c2 (3) d1c1 + d2(c2 - c1) + d3(x - c1 - c2), if c2 x The advantage of this convex flow computation is twofold. It does not require putting thresholds on the arcs a-priori. It is always feasible to have as much flow on a single arc as required. However, the minimum cost flow will avoid the leakage problem because it will be costly to use an erroneous edge to carry the flow from many different patches. 5.1 Fixing the leakage in Isomap As noted earlier, the Isomap method [4] uses the shortest path measurements to estimate a distance matrix M . Afterwards, M is used to find an embedding of the manifold S via MDS. As expected, this method also suffers from the leakage problem as demonstrated in Fig- ure 5. The top-left image in Figure 5 shows our ground truth. In the middle row, we present an embedding of these graphs computed using Isomap which uses the shortest path length as the global distance measure. As illustrated in these figures, even though isomap does a good job in embedding the ground truth when there are no errors, the embedding (or manifold) collapses after we insert the erroneous edges. In contrast, when we use the convex-flow based technique to estimate the distances, we recover the true embedding even in the presence of erroneous edges (Figure 5 bottom row).

Our goal is to recognize human action at a distance, at resolutions where a whole person may be, say, 30 pixels tall. We introduce a novel motion descriptor based on optical flow measurements in a spatiotemporal volume for each stabilized human figure, and an associated similarity measure to be used in a nearest-neighbor framework. Making use of noisy optical flow measurements is the key challenge, which is addressed by treating optical flow not as precise pixel displacements, but rather as a spatial pattern of noisy measurements which are carefully smoothed and aggregated to form our spatiotemporal motion descriptor. To classify the action being performed by a human figure in a query sequence, we retrieve nearest neighbor(s) from a database of stored, annotated video sequences. We can also use these retrieved exemplars to transfer 2D/3D skeletons onto the figures in the query sequence, as well as two forms of data-based action synthesis "do as I do" and "do as I say". Results are demonstrated on ballet, tennis as well as football datasets.

A non-parametric method for texture synthesis is proposed. The texture synthesis process grows a new image outward from an initial seed, one pixel at a time. A Markov random field model is assumed, and the conditional distribution of a pixel given all its neighbors synthesized so far is estimated by querying the sample image and finding all similar neighborhoods. The degree of randomness is controlled by a single perceptually intuitive parameter. The method aims at preserving as much local structure as possible and produces good results for a wide variety of synthetic and real-world textures.