Machine learning interatomic potentials (MLIPs) have become increasingly effective at approximating quantum mechanical calculations at a fraction of the computational cost. However, lower errors on held out test sets do not always translate to improved results on downstream physical property prediction tasks. In this paper, we propose testing MLIPs on their practical ability to conserve energy during molecular dynamic simulations. If passed, improved correlations are found between test errors and their performance on physical property prediction tasks. We identify choices which may lead to models failing this test, and use these observations to improve upon highly-expressive models. The resulting model, eSEN, provides state-of-the-art results on a range of physical property prediction tasks, including materials stability prediction, thermal conductivity prediction, and phonon calculations.

The ability to quickly and accurately compute properties from atomic simulations is critical for advancing a large number of applications in chemistry and materials science including drug discovery, energy storage, and semiconductor manufacturing. To address this need, we present a family of Universal Models for Atoms (UMA), designed to push the frontier of speed, accuracy, and generalization. UMA models are trained on half a billion unique 3D atomic structures (the largest training runs to date) by compiling data across multiple chemical domains, e.g. molecules, materials, and catalysts. We develop empirical scaling laws to help understand how to increase model capacity alongside dataset size to achieve the best accuracy. The UMA small and medium models utilize a novel architectural design we refer to as mixture of linear experts that enables increasing model capacity without sacrificing speed. For example, UMA-medium has 1.4B parameters but only $\sim$50M active parameters per atomic structure. We evaluate UMA models on a diverse set of applications across multiple domains and find that, remarkably, a single model without any fine-tuning can perform similarly or better than specialized models. We are releasing the UMA code, weights, and associated data to accelerate computational workflows and enable the community to build increasingly capable AI models.

We propose fine-tuning large language models for generation of stable materials. While unorthodox, fine-tuning large language models on text-encoded atomistic data is simple to implement yet reliable, with around 90\% of sampled structures obeying physical constraints on atom positions and charges. Using energy above hull calculations from both learned ML potentials and gold-standard DFT calculations, we show that our strongest model (fine-tuned LLaMA-2 70B) can generate materials predicted to be metastable at about twice the rate (49\% vs 28\%) of CDVAE, a competing diffusion model. Because of text prompting's inherent flexibility, our models can simultaneously be used for unconditional generation of stable material, infilling of partial structures and text-conditional generation. Finally, we show that language models' ability to capture key symmetries of crystal structures improves with model scale, suggesting that the biases of pretrained LLMs are surprisingly well-suited for atomistic data.

Foundation models have been transformational in machine learning fields such as natural language processing and computer vision. Similar success in atomic property prediction has been limited due to the challenges of training effective models across multiple chemical domains. To address this, we introduce Joint Multi-domain Pre-training (JMP), a supervised pre-training strategy that simultaneously trains on multiple datasets from different chemical domains, treating each dataset as a unique pre-training task within a multi-task framework. Our combined training dataset consists of $\sim$120M systems from OC20, OC22, ANI-1x, and Transition-1x. We evaluate performance and generalization by fine-tuning over a diverse set of downstream tasks and datasets including: QM9, rMD17, MatBench, QMOF, SPICE, and MD22. JMP demonstrates an average improvement of 59% over training from scratch and matches or sets state-of-the-art on 34 out of 40 tasks. Our work highlights the potential of pre-training strategies that utilize diverse data to advance property prediction across chemical domains, especially for low-data tasks.

Graph neural networks that model 3D data, such as point clouds or atoms, are typically desired to be $SO(3)$ equivariant, i.e., equivariant to 3D rotations. Unfortunately equivariant convolutions, which are a fundamental operation for equivariant networks, increase significantly in computational complexity as higher-order tensors are used. In this paper, we address this issue by reducing the $SO(3)$ convolutions or tensor products to mathematically equivalent convolutions in $SO(2)$ . This is accomplished by aligning the node embeddings' primary axis with the edge vectors, which sparsifies the tensor product and reduces the computational complexity from $O(L^6)$ to $O(L^3)$, where $L$ is the degree of the representation. We demonstrate the potential implications of this improvement by proposing the Equivariant Spherical Channel Network (eSCN), a graph neural network utilizing our novel approach to equivariant convolutions, which achieves state-of-the-art results on the large-scale OC-20 and OC-22 datasets.

Modeling the energy and forces of atomic systems is a fundamental problem in computational chemistry with the potential to help address many of the world’s most pressing problems, including those related to energy scarcity and climate change. These calculations are traditionally performed using Density Functional Theory, which is computationally very expensive. Machine learning has the potential to dramatically improve the efficiency of these calculations from days or hours to seconds.We propose the Spherical Channel Network (SCN) to model atomic energies and forces. The SCN is a graph neural network where nodes represent atoms and edges their neighboring atoms. The atom embeddings are a set of spherical functions, called spherical channels, represented using spherical harmonics. We demonstrate, that by rotating the embeddings based on the 3D edge orientation, more information may be utilized while maintaining the rotational equivariance of the messages. While equivariance is a desirable property, we find that by relaxing this constraint in both message passing and aggregation, improved accuracy may be achieved. We demonstrate state-of-the-art results on the large-scale Open Catalyst 2020 dataset in both energy and force prediction for numerous tasks and metrics.

We introduce the GANsformer, a novel and efficient type of transformer, and explore it for the task of visual generative modeling. The network employs a bipartite structure that enables long-range interactions across the image, while maintaining computation of linear efficiency, that can readily scale to high-resolution synthesis. It iteratively propagates information from a set of latent variables to the evolving visual features and vice versa, to support the refinement of each in light of the other and encourage the emergence of compositional representations of objects and scenes. In contrast to the classic transformer architecture, it utilizes multiplicative integration that allows flexible region-based modulation, and can thus be seen as a generalization of the successful StyleGAN network. We demonstrate the model’s strength and robustness through a careful evaluation over a range of datasets, from simulated multi-object environments to rich real-world indoor and outdoor scenes, showing it achieves state-of-the-art results in terms of image quality and diversity, while enjoying fast learning and better data-efficiency. Further qualitative and quantitative experiments offer us an insight into the model’s inner workings, revealing improved interpretability and stronger disentanglement, and illustrating the benefits and efficacy of our approach. An implementation of the model is available at https://github.com/dorarad/gansformer.

We introduce the GANformer2 model, an iterative object-oriented transformer, explored for the task of generative modeling. The network incorporates strong and explicit structural priors, to reflect the compositional nature of visual scenes, and synthesizes images through a sequential process. It operates in two stages: a fast and lightweight planning phase, where we draft a high-level scene layout, followed by an attention-based execution phase, where the layout is being refined, evolving into a rich and detailed picture. Our model moves away from conventional black-box GAN architectures that feature a flat and monolithic latent space towards a transparent design that encourages efficiency, controllability and interpretability. We demonstrate GANformer2's strengths and qualities through a careful evaluation over a range of datasets, from multi-object CLEVR scenes to the challenging COCO images, showing it successfully achieves state-of-the-art performance in terms of visual quality, diversity and consistency. Further experiments demonstrate the model's disentanglement and provide a deeper insight into its generative process, as it proceeds step-by-step from a rough initial sketch, to a detailed layout that accounts for objects' depths and dependencies, and up to the final high-resolution depiction of vibrant and intricate real-world scenes. See https://github.com/dorarad/gansformer for model implementation.

Magnetic Resonance Image (MRI) acquisition is an inherently slow process which has spurred the development of two different acceleration methods: acquiring multiple correlated samples simultaneously (parallel imaging) and acquiring fewer samples than necessary for traditional signal processing methods (compressed sensing). Both methods provide complementary approaches to accelerating MRI acquisition. In this paper, we present a novel method to integrate traditional parallel imaging methods into deep neural networks that is able to generate high quality reconstructions even for high acceleration factors. The proposed method, called GrappaNet, performs progressive reconstruction by first mapping the reconstruction problem to a simpler one that can be solved by a traditional parallel imaging methods using a neural network, followed by an application of a parallel imaging method, and finally fine-tuning the output with another neural network. The entire network can be trained end-to-end. We present experimental results on the recently released fastMRI dataset and show that GrappaNet can generate higher quality reconstructions than competing methods for both 4× and 8× acceleration.

Research on 2D and 3D generative models typically focuses on the final artifact being created, e.g., an image or a 3D structure. Unlike 2D image generation, the generation of 3D objects in the real world is commonly constrained by the process and order in which the object is constructed. For instance, gravity needs to be taken into account when building a block tower. In this paper, we explore the prediction of ordered actions to construct 3D objects. Instead of predicting actions based on physical constraints, we propose learning through observing human actions. To enable large-scale data collection, we use the Minecraft1 environment. We introduce 3D-Craft, a new dataset of 2,500 Minecraft houses each built by human players sequentially from scratch. To learn from these human action sequences, we propose an order-aware 3D generative model called VoxelCNN. In contrast to other 3D generative models which either have no explicit order (e.g. holistic generation with 3DGAN [35]), or follow a simple heuristic order (e.g. raster-scan), VoxelCNN is trained to imitate human building order with spatial awareness. We also transferred the order to other dataset such as ShapeNet[10]. The 3D-Craft dataset, models, and benchmark system will be made publicly available, which may inspire new directions for future research exploration. https://github.com/facebookresearch/VoxelCNN.

The AlphaGo, AlphaGo Zero, and AlphaZero series of algorithms are remarkable demonstrations of deep reinforcement learning’s capabilities, achieving superhuman performance in the complex game of Go with progressively increasing autonomy. However, many obstacles remain in the understanding of and usability of these promising approaches by the research community. Toward elucidating unresolved mysteries and facilitating future research, we propose ELF OpenGo, an open-source reimplementation of the AlphaZero algorithm. ELF OpenGo is the first open-source Go AI to convincingly demonstrate superhuman performance with a perfect (20:0) record against global top professionals. We apply ELF OpenGo to conduct extensive ablation studies, and to identify and analyze numerous interesting phenomena in both the model training and in the gameplay inference procedures. Our code, models, selfplay datasets, and auxiliary data are publicly available.

When building artificial intelligence systems that can reason and answer questions about visual data, we need diagnostic tests to analyze our progress and discover short-comings. Existing benchmarks for visual question answering can help, but have strong biases that models can exploit to correctly answer questions without reasoning. They also conflate multiple sources of error, making it hard to pinpoint model weaknesses. We present a diagnostic dataset that tests a range of visual reasoning abilities. It contains minimal biases and has detailed annotations describing the kind of reasoning each question requires. We use this dataset to analyze a variety of modern visual reasoning systems, providing novel insights into their abilities and limitations.

Existing methods for visual reasoning attempt to directly map inputs to outputs using black-box architectures without explicitly modeling the underlying reasoning processes. As a result, these black-box models often learn to exploit biases in the data rather than learning to perform visual reasoning. Inspired by module networks, this paper proposes a model for visual reasoning that consists of a program generator that constructs an explicit representation of the reasoning process to be performed, and an execution engine that executes the resulting program to produce an answer. Both the program generator and the execution engine are implemented by neural networks, and are trained using a combination of backpropagation and REINFORCE. Using the CLEVR benchmark for visual reasoning, we show that our model significantly outperforms strong baselines and generalizes better in a variety of settings.

Interactive agents are becoming increasingly common in many application domains, such as education, healthcare and personal assistance. The success of such embodied agents relies on their ability to have sustained engagement with their human users. Such engagement requires agents to be socially intelligent, equipped with the ability to understand and reciprocate both verbal and non-verbal cues. While there has been tremendous progress in verbal communication, mostly driven by the success of speech recognition and question-answering, teaching agents to appropriately react to facial expressions has received less attention. In this paper, we focus on non-verbal facial cues for face-to-face communication between a user and an embodied agent. We propose a method that automatically learns to update the agent's facial expressions based on the user's expressions. We adopt a learning scheme and train a deep neural network on hundreds of videos, containing pairs of people engaging in a conversation, and without external human supervision. Our experimental results show the efficacy of our model in sustained long-term prediction of the agent's facial landmarks. We present comparative results showing that our model significantly outperforms baseline approaches and provide insightful human studies to better understand our model's qualitative performance. We release our dataset to further encourage research in this field.

In this paper, we propose ELF, an Extensive, Lightweight and Flexible platform for fundamental reinforcement learning research. Using ELF, we implement a highly customizable real-time strategy (RTS) engine with three game environments (Mini-RTS, Capture the Flag and Tower Defense). Mini-RTS, as a miniature version of StarCraft, captures key game dynamics and runs at 165K frame-per-second (FPS) on a laptop. When coupled with modern reinforcement learning methods, the system can train a full-game bot against built-in AIs end-to-end in one day with 6 CPUs and 1 GPU. In addition, our platform is flexible in terms of environment-agent communication topologies, choices of RL methods, changes in game parameters, and can host existing C/C++-based game environments like ALE. Using ELF, we thoroughly explore training parameters and show that a network with Leaky ReLU and Batch Normalization coupled with long-horizon training and progressive curriculum beats the rule-based built-in AI more than 70% of the time in the full game of Mini-RTS. Strong performance is also achieved on the other two games. In game replays, we show our agents learn interesting strategies. ELF, along with its RL platform, is open-sourced at https://github.com/facebookresearch/ELF.

It is well known that contextual and multi-scale representations are important for accurate visual recognition. In this paper we present the Inside-Outside Net (ION), an object detector that exploits information both inside and outside the region of interest. Contextual information outside the region of interest is integrated using spatial recurrent neural networks. Inside, we use skip pooling to extract information at multiple scales and levels of abstraction. Through extensive experiments we evaluate the design space and provide readers with an overview of what tricks of the trade are important. ION improves state-of-the-art on PASCAL VOC 2012 object detection from 73.9% to 77.9% mAP. On the new and more challenging MS COCO dataset, we improve state-of-the-art from 19.7% to 33.1% mAP. In the 2015 MS COCO Detection Challenge, our ION model won "Best Student Entry" and finished 3rd place overall. As intuition suggests, our detection results provide strong evidence that context and multi-scale representations improve small object detection.

Humor is an integral part of human lives. Despite being tremendously impactful, it is perhaps surprising that we do not have a detailed understanding of humor yet. As interactions between humans and AI systems increase, it is imperative that these systems are taught to understand subtleties of human expressions such as humor. In this work, we are interested in the question – what content in a scene causes it to be funny? As a first step towards understanding visual humor, we analyze the humor manifested in abstract scenes and design computational models for them. We collect two datasets of abstract scenes that facilitate the study of humor at both the scene-level and the object-level. We analyze the funny scenes and explore the different types of humor depicted in them via human studies. We model two tasks that we believe demonstrate an understanding of some aspects of visual humor. The tasks involve predicting the funniness of a scene and altering the funniness of a scene. We show that our models perform well quantitatively, and qualitatively through human studies. Our datasets are publicly available.

When human annotators are given a choice about what to label in an image, they apply their own subjective judgments on what to ignore and what to mention. We refer to these noisy "human-centric" annotations as exhibiting human reporting bias. Examples of such annotations include image tags and keywords found on photo sharing sites, or in datasets containing image captions. In this paper, we use these noisy annotations for learning visually correct image classifiers. Such annotations do not use consistent vocabulary, and miss a significant amount of the information present in an image, however, we demonstrate that the noise in these annotations exhibits structure and can be modeled. We propose an algorithm to decouple the human reporting bias from the correct visually grounded labels. Our results are highly interpretable for reporting "what's in the image" versus "what's worth saying." We demonstrate the algorithm's efficacy along a variety of metrics and datasets, including MS COCO and Yahoo Flickr 100M.We show significant improvements over traditional algorithms for both image classification and image captioning, doubling the performance of existing methods in some cases.

We conduct large-scale studies on ‘human attention’ in Visual Question Answering (VQA) to understand where humans choose to look to answer questions about images. We design and test multiple game-inspired novel attention-annotation interfaces that require the subject to sharpen regions of a blurred image to answer a question. Thus, we introduce the VQA-HAT (Human ATtention) dataset. We evaluate attention maps generated by state-of-the-art VQA models against human attention both qualitatively (via visualizations) and quantitatively (via rank-order correlation). Our experiments show that current attention models in VQA do not seem to be looking at the same regions as humans. Finally, we train VQA models with explicit attention supervision, and find that it improves VQA performance.

Ting-Hao Kenneth Huang, Francis Ferraro, Nasrin Mostafazadeh, Ishan Misra, Aishwarya Agrawal, Jacob Devlin, Ross Girshick, Xiaodong He, Pushmeet Kohli, Dhruv Batra, C. Lawrence Zitnick, Devi Parikh, Lucy Vanderwende, Michel Galley, Margaret Mitchell. Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2016.

Relating visual information to its linguistic semantic meaning remains an open and challenging area of research. The semantic meaning of images depends on the presence of objects, their attributes and their relations to other objects. But precisely characterizing this dependence requires extracting complex visual information from an image, which is in general a difficult and yet unsolved problem. In this paper, we propose studying semantic information in abstract images created from collections of clip art. Abstract images provide several advantages over real images. They allow for the direct study of how to infer high-level semantic information, since they remove the reliance on noisy low-level object, attribute and relation detectors, or the tedious hand-labeling of real images. Importantly, abstract images also allow the ability to generate sets of semantically similar scenes. Finding analogous sets of real images that are semantically similar would be nearly impossible. We create 1,002 sets of 10 semantically similar abstract images with corresponding written descriptions. We thoroughly analyze this dataset to discover semantically important features, the relations of words to visual features and methods for measuring semantic similarity. Finally, we study the relation between the saliency and memorability of objects and their semantic importance.

In this paper we explore the bi-directional mapping between images and their sentence-based descriptions. Critical to our approach is a recurrent neural network that attempts to dynamically build a visual representation of the scene as a caption is being generated or read. The representation automatically learns to remember long-term visual concepts. Our model is capable of both generating novel captions given an image, and reconstructing visual features given an image description. We evaluate our approach on several tasks. These include sentence generation, sentence retrieval and image retrieval. State-of-the-art results are shown for the task of generating novel image descriptions. When compared to human generated captions, our automatically generated captions are equal to or preferred by humans 21.0% of the time. Results are better than or comparable to state-of-the-art results on the image and sentence retrieval tasks for methods using similar visual features.

This paper presents a novel approach for automatically generating image descriptions: visual detectors, language models, and multimodal similarity models learnt directly from a dataset of image captions. We use multiple instance learning to train visual detectors for words that commonly occur in captions, including many different parts of speech such as nouns, verbs, and adjectives. The word detector outputs serve as conditional inputs to a maximum-entropy language model. The language model learns from a set of over 400,000 image descriptions to capture the statistics of word usage. We capture global semantics by re-ranking caption candidates using sentence-level features and a deep multimodal similarity model. Our system is state-of-the-art on the official Microsoft COCO benchmark, producing a BLEU-4 score of 29.1%. When human judges compare the system captions to ones written by other people on our held-out test set, the system captions have equal or better quality 34% of the time.

Automatically describing an image with a sentence is a long-standing challenge in computer vision and natural language processing. Due to recent progress in object detection, attribute classification, action recognition, etc., there is renewed interest in this area. However, evaluating the quality of descriptions has proven to be challenging. We propose a novel paradigm for evaluating image descriptions that uses human consensus. This paradigm consists of three main parts: a new triplet-based method of collecting human annotations to measure consensus, a new automated metric that captures consensus, and two new datasets: PASCAL-50S and ABSTRACT-50S that contain 50 sentences describing each image. Our simple metric captures human judgment of consensus better than existing metrics across sentences generated by various sources. We also evaluate five state-of-the-art image description approaches using this new protocol and provide a benchmark for future comparisons. A version of CIDEr named CIDEr-D is available as a part of MS COCO evaluation server to enable systematic evaluation and benchmarking.

We propose the task of free-form and open-ended Visual Question Answering (VQA). Given an image and a natural language question about the image, the task is to provide an accurate natural language answer. Mirroring real-world scenarios, such as helping the visually impaired, both the questions and answers are open-ended. Visual questions selectively target different areas of an image, including background details and underlying context. As a result, a system that succeeds at VQA typically needs a more detailed understanding of the image and complex reasoning than a system producing generic image captions. Moreover, VQA is amenable to automatic evaluation, since many open-ended answers contain only a few words or a closed set of answers that can be provided in a multiple-choice format. We provide a dataset containing ~0.25M images, ~0.76M questions, and ~10M answers (www.visualqa.org), and discuss the information it provides. Numerous baselines for VQA are provided and compared with human performance.

Common sense is essential for building intelligent machines. While some commonsense knowledge is explicitly stated in human-generated text and can be learnt by mining the web, much of it is unwritten. It is often unnecessary and even unnatural to write about commonsense facts. While unwritten, this commonsense knowledge is not unseen! The visual world around us is full of structure modeled by commonsense knowledge. Can machines learn common sense simply by observing our visual world? Unfortunately, this requires automatic and accurate detection of objects, their attributes, poses, and interactions between objects, which remain challenging problems. Our key insight is that while visual common sense is depicted in visual content, it is the semantic features that are relevant and not low-level pixel information. In other words, photorealism is not necessary to learn common sense. We explore the use of human-generated abstract scenes made from clipart for learning common sense. In particular, we reason about the plausibility of an interaction or relation between a pair of nouns by measuring the similarity of the relation and nouns with other relations and nouns we have seen in abstract scenes. We show that the commonsense knowledge we learn is complementary to what can be learnt from sources of text.

In this paper, we study the problem of answering visual analogy questions. These questions take the form of image A is to image B as image C is to what. Answering these questions entails discovering the mapping from image A to image B and then extending the mapping to image C and searching for the image D such that the relation from A to B holds for C to D. We pose this problem as learning an embedding that encourages pairs of analogous images with similar transformations to be close together using convolutional neural networks with a quadruple Siamese architecture. We introduce a dataset of visual analogy questions in natural images, and show first results of its kind on solving analogy questions on natural images.

Edge detection is a critical component of many vision systems, including object detectors and image segmentation algorithms. Patches of edges exhibit well-known forms of local structure, such as straight lines or T-junctions. In this paper we take advantage of the structure present in local image patches to learn both an accurate and computationally efficient edge detector. We formulate the problem of predicting local edge masks in a structured learning framework applied to random decision forests. Our novel approach to learning decision trees robustly maps the structured labels to a discrete space on which standard information gain measures may be evaluated. The result is an approach that obtains realtime performance that is orders of magnitude faster than many competing state-of-the-art approaches, while also achieving state-of-the-art edge detection results on the BSDS500 Segmentation dataset and NYU Depth dataset. Finally, we show the potential of our approach as a general purpose edge detector by showing our learned edge models generalize well across datasets.

Given a static scene, a human can trivially enumerate the myriad of things that can happen next and characterize the relative likelihood of each. In the process, we make use of enormous amounts of commonsense knowledge about how the world works. In this paper, we investigate learning this commonsense knowledge from data. To overcome a lack of densely annotated spatiotemporal data, we learn from sequences of abstract images gathered using crowd-sourcing. The abstract scenes provide both object location and attribute information. We demonstrate qualitatively and quantitatively that our models produce plausible scene predictions on both the abstract images, as well as natural images taken from the Internet.

Several popular and effective object detectors separately model intra-class variations arising from deformations and appearance changes. This reduces model complexity while enabling the detection of objects across changes in view- point, object pose, etc. The Deformable Part Model (DPM) is perhaps the most successful such model to date. A common assumption is that the exponential number of templates enabled by a DPM is critical to its success. In this paper, we show the counter-intuitive result that it is possible to achieve similar accuracy using a small dictionary of deformations. Each component in our model is represented by a single HOG template and a dictionary of flow fields that determine the deformations the template may undergo. While the number of candidate deformations is dramatically fewer than that for a DPM, the deformed templates tend to be plausible and interpretable. In addition, we discover that the set of deformation bases is actually transferable across object categories and that learning shared bases across similar categories can boost accuracy.

We present a new approach for recognizing the make and model of a car from a single image. While most previous methods are restricted to fixed or limited viewpoints, our system is able to verify a car's make and model from an arbitrary view. Our model consists of 3D space curves obtained by backprojecting image curves onto silhouette-based visual hulls and then refining them using three-view curve matching. We also build an appearance model of taillights which is used as an additional cue. Our approach is able to verify the exact make and model of a car over a wide range of viewpoints and background clutter.

We propose a novel approach to both learning and detecting local contour-based representations for mid-level features. Our features, called sketch tokens, are learned using supervised mid-level information in the form of hand drawn contours in images. Patches of human generated contours are clustered to form sketch token classes and a random forest classifier is used for efficient detection in novel images. We demonstrate our approach on both top-down and bottom-up tasks. We show state-of-the-art results on the top-down task of contour detection while being over 200x faster than competing methods. We also achieve large improvements in detection accuracy for the bottom-up tasks of pedestrian and object detection as measured on INRIA and PASCAL, respectively. These gains are due to the complementary information provided by sketch tokens to low-level features such as gradient histograms.

We describe novel but simple motion features for the problem of detecting objects in video sequences. Previous approaches either compute optical flow or temporal differences on video frame pairs with various assumptions about stabilization. We describe a combined approach that uses coarse-scale flow and fine-scale temporal difference features. Our approach performs weak motion stabilization by factoring out camera motion and coarse object motion while preserving nonrigid motions that serve as useful cues for recognition. We show results for pedestrian detection and human pose estimation in video sequences, achieving state-of-the-art results in both. In particular, given a fixed detection rate our method achieves a five-fold reduction in false positives over prior art on the Caltech Pedestrian benchmark. Finally, we perform extensive diagnostic experiments to reveal what aspects of our system are crucial for good performance. Proper stabilization, long time-scale features, and proper normalization are all critical.

Relating visual information to its linguistic semantic meaning remains an open and challenging area of research. The semantic meaning of images depends on the presence of objects, their attributes and their relations to other objects. But precisely characterizing this dependence requires extracting complex visual information from an image, which is in general a difficult and yet unsolved problem. In this paper, we propose studying semantic information in abstract images created from collections of clip art. Abstract images provide several advantages. They allow for the direct study of how to infer high-level semantic information, since they remove the reliance on noisy low-level object, attribute and relation detectors, or the tedious hand-labeling of images. Importantly, abstract images also allow the ability to generate sets of semantically similar scenes. Finding analogous sets of semantically similar real images would be nearly impossible. We create 1,002 sets of 10 semantically similar abstract scenes with corresponding written descriptions. We thoroughly analyze this dataset to discover semantically important features, the relations of words to visual features and methods for measuring semantic similarity.

Edge detection is a critical component of many vision systems, including object detectors and image segmentation algorithms. Patches of edges exhibit well-known forms of local structure, such as straight lines or T-junctions. In this paper we take advantage of the structure present in local image patches to learn both an accurate and computationally efficient edge detector. We formulate the problem of predicting local edge masks in a structured learning framework applied to random decision forests. Our novel approach to learning decision trees robustly maps the structured labels to a discrete space on which standard information gain measures may be evaluated. The result is an approach that obtains real time performance that is orders of magnitude faster than many competing state-of-the-art approaches, while also achieving state-of-the-art edge detection results on the BSDS500 Segmentation dataset and NYU Depth dataset. Finally, we show the potential of our approach as a general purpose edge detector by showing our learned edge models generalize well across datasets.

Sentences that describe visual scenes contain a wide variety of information pertaining to the presence of objects, their attributes and their spatial relations. In this paper we learn the visual features that correspond to semantic phrases derived from sentences. Specifically, we extract predicate tuples that contain two nouns and a relation. The relation may take several forms, such as a verb, preposition, adjective or their combination. We model a scene using a Conditional Random Field (CRF) formulation where each node corresponds to an object, and the edges to their relations. We determine the potentials of the CRF using the tuples extracted from the sentences. We generate novel scenes depicting the sentences' visual meaning by sampling from the CRF. The CRF is also used to score a set of scenes for a text-based image retrieval task. Our results show we can generate (retrieve) scenes that convey the desired semantic meaning, even when scenes (queries) are described by multiple sentences. Significant improvement is found over several baseline approaches.

In this paper, we propose a general purpose approach to handwriting beautification using online input from a stylus. Given a sample of writings, drawings, or sketches from the same user, our method improves a user's strokes in real-time as they are drawn. Our approach relies on one main insight. The appearance of the average of multiple instances of the same written word or shape is better than most of the individual instances. We utilize this observation using a two-stage approach. First, we propose an efficient real-time method for finding matching sets of stroke samples called tokens in a potentially large database of writings from a user. Second, we refine the user's most recently written strokes by averaging them with the matching tokens. Our approach works without handwriting recognition, and does not require a database of predefined letters, words, or shapes. Our results show improved results for a wide range of writing styles and drawings.

This paper addresses the problem of image alignment using direct intensity-based methods for affine and homography transformations. Direct methods often employ scale-space smoothing (Gaussian blur) of the images to avoid local minima. Although, it is known that the isotropic blur used is not optimal for some motion models, the correct blur kernels have not been rigorously derived for motion models beyond translations. In this work, we derive blur kernels that result from smoothing the alignment objective function for some common motion models such as affine and homography. We show the derived kernels remove poor local minima and reach lower energy solutions in practice.

Many cues have been proposed for contour detection or image segmentation. These include low-level image gradients to high-level information such as the identity of the objects in the scene or 3D depth understanding. While state-of-the-art approaches have been incorporating more cues, the relative importance of the cues is unclear. In this paper, we examine the relative importance of low-, mid- and high-level cues to gain a better understanding of their role in detecting object contours in an image. To accomplish this task, we conduct numerous human studies and compare their performance to several popular segmentation and contour detection machine approaches. Our findings suggest that the current state-of-the-art contour detection algorithms perform as well as humans using low-level cues. We also find evidence that the recognition of objects, but not occlusion information, leads to improved human performance. Moreover, when objects are recognized by humans, their contour detection performance increases over current machine algorithms. Finally, mid-level cues appear to offer a larger performance boost than high-level cues such as recognition.

The restoration of a blurry or noisy image is commonly performed with a MAP estimator, which maximizes a posterior probability to reconstruct a clean image from a degraded image. A MAP estimator, when used with a sparse gradient image prior, reconstructs piecewise smooth images and typically removes textures that are important for visual realism. We present an alternative deconvolution method called iterative distribution reweighting (IDR) which imposes a global constraint on gradients so that a reconstructed image should have a gradient distribution similar to a reference distribution. In natural images, a reference distribution not only varies from one image to another, but also within an image depending on texture. We estimate a reference distribution directly from an input image for each texture segment. Our algorithm is able to restore rich mid-frequency textures. A large-scale user study supports the conclusion that our algorithm improves the visual realism of reconstructed images compared to those of MAP estimators.

Typically, object recognition is performed based solely on the appearance of the object. However, relevant information also exists in the scene surrounding the object. In this paper, we explore the roles that appearance and contextual information play in object recognition. Through machine experiments and human studies, we show that the importance of contextual information varies with the quality of the appearance information, such as an image's resolution. Our machine experiments explicitly model context between object categories through the use of relative location and relative scale, in addition to co-occurrence. With the use of our context model, our algorithm achieves state-of-the-art performance on the MSRC and Corel data sets. We perform recognition tests for machines and human subjects on low and high resolution images, which vary significantly in the amount of appearance information present, using just the object appearance information, the combination of appearance and context, as well as just context without object appearance information (blind recognition). We also explore the impact of the different sources of context (co-occurrence, relative-location, and relative-scale). We find that the importance of different types of contextual information varies significantly across data sets such as MSRC and PASCAL.

Detecting people remains a popular and challenging problem in computer vision. In this paper, we analyze parts-based models for person detection to determine which components of their pipeline could benefit the most if improved. We accomplish this task by studying numerous detectors formed from combinations of components performed by human subjects and machines. The parts-based model we study can be roughly broken into four components: feature detection, part detection, spatial part scoring and contextual reasoning including non-maximal suppression. Our experiments conclude that part detection is the weakest link for challenging person detection datasets. Non-maximal suppression and context can also significantly boost performance. However, the use of human or machine spatial models does not significantly or consistently affect detection accuracy.

In this paper, we describe an interest point detector using edge foci. Unlike traditional detectors that compute interest points directly from image intensities, we use normalized intensity edges and their orientations. We hypothesize that detectors based on the presence of oriented edges are more robust to non-linear lighting variations and background clutter than intensity based techniques. Specifically, we detect edge foci, which are points in the image that are roughly equidistant from edges with orientations perpendicular to the point. The scale of the interest point is defined by the distance between the edge foci and the edges. We quantify the performance of our detector using the interest point's repeatability, uniformity of spatial distribution, and the uniqueness of the resulting descriptors. Results are found using traditional datasets and new datasets with challenging non-linear lighting variations and occlusions.

We present ShadowDraw, a system for guiding the freeform drawing of objects. As the user draws, ShadowDraw dynamically updates a shadow image underlying the user's strokes. The shadows are suggestive of object contours that guide the user as they continue drawing. This paradigm is similar to tracing, with two major differences. First, we do not provide a single image from which the user can trace; rather ShadowDraw automatically blends relevant images from a large database to construct the shadows. Second, the system dynamically adapts to the user's drawings in real-time and produces suggestions accordingly. ShadowDraw works by efficiently matching local edge patches between the query, constructed from the current drawing, and a database of images. A hashing technique enforces both local and global similarity and provides sufficient speed for interactive feedback. Shadows are created by aggregating the edge maps from the best database matches, spatially weighted by their match scores. We test our approach with human subjects and show comparisons between the drawings that were produced with and without the system. The results show that our system produces more realistically proportioned line drawings.

In image restoration tasks, a heavy-tailed gradient distribution of natural images has been extensively exploited as an image prior. Most image restoration algorithms impose a sparse gradient prior on the whole image, reconstructing an image with piecewise smooth characteristics. While the sparse gradient prior removes ringing and noise artifacts, it also tends to remove mid-frequency textures, degrading the visual quality. We can attribute such degradations to imposing an incorrect image prior. The gradient profile in fractal-like textures, such as trees, is close to a Gaussian distribution, and small gradients from such regions are severely penalized by the sparse gradient prior. To address this issue, we introduce an image restoration algorithm that adapts the image prior to the underlying texture. We adapt the prior to both low-level local structures as well as mid-level textural characteristics. Improvements in visual quality is demonstrated on deconvolution and denoising tasks.

There are many computer vision algorithms developed for visual (scene and object) recognition. Some systems focus on involved learning algorithms, some leverage millions of training images, and some systems focus on modeling relevant information (features) with the goal of effective recognition. However, none of these systems come close to human capabilities. If we study human responses on similar problems we could gain insight into which of the three factors (1) learning algorithm (2) amount of training data and (3) features is critical to humans' superior performance. In this work we take a small step towards this goal by performing a series of human studies and machine experiments. We find no evidence that human pattern matching algorithms are better than standard machine learning algorithms. Moreover, we find that humans don't leverage increased amounts of training data. Through statistical analysis on the machine experiments and supporting human studies, we find that the main factor impacting accuracies is the choice of features.

We present an optimization framework for exploring gradient-domain solutions for image and video processing. The proposed framework unifies many of the key ideas in the gradient-domain literature under a single optimization formulation. Our hope is that this generalized framework will allow the reader to quickly gain a general understanding of the field and contribute new ideas of their own. We propose a novel metric for measuring local gradient saliency that identifies salient gradients that give rise to long, coherent edges, even when the individual gradients are faint. We present a general weighting scheme for gradient constraints that improves the visual appearance of results. We also provide a solution for applying gradient-domain filters to videos and video streams in a coherent manner. Finally, we demonstrate the utility of our formulation in creating effective yet simple to implement solutions for various image-processing tasks. To exercise our formulation we have created a new saliency-based sharpen filter and a pseudo image-relighting application. We also revisit and improve upon previously defined filters such as nonphotorealistic rendering, image deblocking, and sparse data interpolation over images (e.g., colorization using optimization).

We present a deblurring algorithm that uses a hardware attachment coupled with a natural image prior to deblur images from consumer cameras. Our approach uses a combination of inexpensive gyroscopes and accelerometers in an energy optimization framework to estimate a blur function from the camera's acceleration and angular velocity during an exposure. We solve for the camera motion at a high sampling rate during an exposure and infer the latent image using a joint optimization. Our method is completely automatic, handles per-pixel, spatially-varying blur, and out-performs the current leading image-based methods. Our experiments show that it handles large kernels -- up to at least 100 pixels, with a typical size of 30 pixels. We also present a method to perform "ground-truth" measurements of camera motion blur. We use this method to validate our hardware and deconvolution approach. To the best of our knowledge, this is the first work that uses 6 DOF inertial sensors for dense, per-pixel spatially-varying image deblurring and the first work to gather dense ground-truth measurements for camera-shake blur.

We propose an algorithm to cluster video shots by the location in which they were captured. Each shot is represented as a set of keyframes and each keyframe is represented by a histogram of textons. Clustering is performed using an energy-based formulation. We propose an energy function for the clusters that matches the expected distribution of viewpoints in any one location and use the chi-squared distance to measure the similarity of two shots. We also add a temporal prior to model the fact that temporally neighboring shots are more likely to have been captured in the same location. We test our algorithm on both home videos and professionally edited footage (sitcoms). Quantitative results are presented to justify each choice made in the design of our algorithm, as well as comparisons with k-means, connected components, and spectral clustering. 1

Image blur and noise are difficult to avoid in many situations and can often ruin a photograph. We present a novel image deconvolution algorithm that deblurs and denoises an image given a known shift-invariant blur kernel. Our algorithm uses local color statistics derived from the image as a constraint in a unified framework that can be used for deblurring, denoising, and upsampling. A pixel's color is required to be a linear combination of the two most prevalent colors within a neighborhood of the pixel. This two-color prior has two major benefits: it is tuned to the content of the particular image and it serves to decouple edge sharpness from edge strength. Our unified algorithm for deblurring and denoising out-performs previous methods that are specialized for these individual applications. We demonstrate this with both qualitative results and extensive quantitative comparisons that show that we can out-perform previous methods by approximately 1 to 3 DB.

The visual world demonstrates organized spatial patterns, among objects or regions in a scene, object-parts in an object, and low-level features in object-parts. These classes of spatial structures are inherently hierarchical in nature. Although seemingly quite different these spatial patterns are simply manifestations of different levels in a hierarchy. In this work, we present a unified approach to unsupervised learning of hierarchical spatial structures from a collection of images. Ours is a hierarchical rule-based model capturing spatial patterns, where each rule is represented by a star-graph. We propose an unsupervised EM-style algorithm to learn our model from a collection of images. We show that the inference problem of determining the set of learnt rules instantiated in an image is equivalent to finding the minimum-cost Steiner tree in a directed acyclic graph. We evaluate our approach on a diverse set of data sets of object categories, natural outdoor scenes and images from complex street scenes with multiple objects.

Traditionally, object recognition is performed based solely on the appearance of the object. However, relevant information also exists in the scene surrounding the object. As supported by our human studies, this contextual information is necessary for accurate recognition in low resolution images. This scenario with impoverished appearance information, as opposed to using images of higher resolution, provides an appropriate venue for studying the role of context in recognition. In this paper, we explore the role of context for dense scene labeling in small images. Given a segmentation of an image, our algorithm assigns each segment to an object category based on the segment’s appearance and contextual information. We explicitly model context between object categories through the use of relative location and relative scale, in addition to co-occurrence. We perform recognition tests on low and high resolution images, which vary significantly in the amount of appearance information present, using just the object appearance information, the combination of appearance and context, as well as just context without object appearance information (blind recognition). We also perform these tests in human studies and analyze our findings to reveal interesting patterns. With the use of our context model, our algorithm achieves state-of-the-art performance on MSRC and Corel. datasets.

We present an over-segmentation based, dense stereo algorithm that jointly estimates segmentation and depth. For mixed pixels on segment boundaries, the algorithm computes foreground opacity (alpha), as well as color and depth for the foreground and background. We model the scene as a collection of fronto-parallel planar segments in a reference view, and use a generative model for image formation that handles mixed pixels at segment boundaries. Our method iteratively updates the segmentation based on color, depth and shape constraints using MAP estimation. Given a segmentation, the depth estimates are updated using belief propagation. We show that our method is competitive with the state-of-the-art based on the new Middlebury stereo evaluation, and that it overcomes limitations of traditional segmentation based methods while properly handling mixed pixels. Z-keying results show the advantages of combining opacity and depth estimation.

Image denoising algorithms often assume an additive white Gaussian noise (AWGN) process that is independent of the actual RGB values. Such approaches are not fully automatic and cannot effectively remove color noise produced by todays CCD digital camera. In this paper, we propose a unified framework for two tasks: automatic estimation and removal of color noise from a single image using piecewise smooth image models. We introduce the noise level function (NLF), which is a continuous function describing the noise level as a function of image brightness. We then estimate an upper bound of the real noise level function by fitting a lower envelope to the standard deviations of per-segment image variances. For denoising, the chrominance of color noise is significantly removed by projecting pixel values onto a line fit to the RGB values in each segment. Then, a Gaussian conditional random field (GCRF) is constructed to obtain the underlying clean image from the noisy input. Extensive experiments are conducted to test the proposed algorithm, which is shown to outperform state-of-the-art denoising algorithms.

The goal of this work is to find all occurrences of a particular person in a sequence of photographs taken over a short period of time. For identification, we assume each individual’s hair and clothing stays the same throughout the sequence. Even with these assumptions, the task remains challenging as people can move around, change their pose and scale, and partially occlude each other. We propose a two stage method. First, individuals are identified by clustering frontal face detections using color clothing information. Second, a color based pictorial structure model is used to find occurrences of each person in images where their frontal face detection was missed. Two extensions improving the pictorial structure detections are also described. In the first extension, we obtain a better clothing segmentation to improve the accuracy of the clothing color model. In the second extension, we simultaneously consider multiple detection hypotheses of all people potentially present in the shot. Our results show that people can be re-detected in images where they do not face the camera. Results are presented on several sequences from a personal photo collection.

Recently, the generative modeling approach to video segmentation has been gaining popularity in the computer vision community. For example, the flexible sprites framework has been studied in, among other references, [11,13,14,24]. In general, detailed generative models are vulnerable to intractability of inference and local minima problems when approximations are made (see, e.g., [25]). Recent approaches to dealing with these problems focused on inference techniques for increasingly more expressive models. Simpler models, on the other hand, while less precise, are often not just faster, but less prone to local minima. In addition, while many different models may be based on similar hidden variables, some models may be more amenable to inference of some of the shared variables, while other models lead to efficient and accurate inference of other components of the hierarchical data description. In this paper, we empirically illustrate that forcing multiple models to share the posterior distribution leads to inference less prone to local minima. We define a set of key hidden variables that describe aspects of the data that we care about. The relationships among these key variables are defined through multiple conditional distribution models on the same pairs of variables, controlled by switch variables. The posterior distribution over the key hidden variables is shared, and inference of the switch variables serves as a mechanism for combinatorial model selection. The key observation here is that while the most expressive model often ends up a winner by the end of the iterative learning of model parameters, early iterations are dominated by simpler model components, and upon convergence, the free energy is lower than the ones reached by switching on all the most complex components from the beginning of the learning. We illustrate the performance of this approach on the unsupervised video segmentation task.

Most algorithms for 3D reconstruction from images use cost functions based on SSD, which assume that the surfaces being reconstructed are visible to all cameras. This makes it difficult to reconstruct objects which are partially occluded. Recently, researchers working with large camera arrays have shown it is possible to "see through" occlusions using a technique called synthetic aperture focusing. This suggests that we can design alternative cost functions that are robust to occlusions using synthetic apertures. Our paper explores this design space. We compare classical shape from stereo with shape from synthetic aperture focus. We also describe two variants of multi-view stereo based on color medians and entropy that increase robustness to occlusions. We present an experimental comparison of these cost functions on complex light fields, measuring their accuracy against the amount of occlusion.

In this paper, we propose a method for jointly computing optical flow and segmenting video while accounting for mixed pixels (matting). Our method is based on statistical modeling of an image pair using constraints on appearance and motion. Segments are viewed as overlapping regions with fractional (/spl alpha/) contributions. Bidirectional motion is estimated based on spatial coherence and similarity of segment colors. Our model is extended to video by chaining the pairwise models to produce a joint probability distribution to be maximized. To make the problem more tractable, we factorize the posterior distribution and iteratively minimize its parts. We demonstrate our method on frame interpolation.

The ability to interactively control viewpoint while watching a video is an exciting application of image-based rendering. The goal of our work is to render dynamic scenes with interactive viewpoint control using a relatively small number of video cameras. In this paper, we show how high-quality video-based rendering of dynamic scenes can be accomplished using multiple synchronized video streams combined with novel image-based modeling and rendering algorithms. Once these video streams have been processed, we can synthesize any intermediate view between cameras at any time, with the potential for space-time manipulation.In our approach, we first use a novel color segmentation-based stereo algorithm to generate high-quality photoconsistent correspondences across all camera views. Mattes for areas near depth discontinuities are then automatically extracted to reduce artifacts during view synthesis. Finally, a novel temporal two-layer compressed representation that handles matting is developed for rendering at interactive rates.

Presents a stereo algorithm for obtaining disparity maps with occlusion explicitly detected. To produce smooth and detailed disparity maps, two assumptions that were originally proposed by Marr and Poggio (1976, 1979) are adopted: uniqueness and continuity. That is, the disparity maps have a unique value per pixel and are continuous almost everywhere. These assumptions are enforced within a three-dimensional array of match values in disparity space. Each match value corresponds to a pixel in an image and a disparity relative to another image. An iterative algorithm updates the match values by diffusing support among neighboring values and inhibiting others along similar lines of sight. By applying the uniqueness assumption, occluded regions can be explicitly identified. To demonstrate the effectiveness of the algorithm, we present the processing results from synthetic and real image pairs, including ones with ground-truth values for quantitative comparison with other methods.

We describe our implementation of a parallel depth recovery scheme for a four-camera multibaseline stereo in a convergent configuration. Our system is capable of image capture at video rate. This is critical in applications that require three-dimensional tracking. We obtain dense stereo depth data by projecting a light pattern of frequency modulated sinusoidally varying intensity onto the scene, thus increasing the local discriminability at each pixel and facilitating matches. In addition, we make most of the camera view areas by converging them at a volume of interest. Results show that we are able to extract stereo depth data that are, on the average, less than 1 mm in error at distances between 1.5 to 3.5 m away from the cameras.>