Abhinav Gupta

Suraj Nair, Aravind Rajeswaran, Vikash Kumar et al. · stanford

We study how visual representations pre-trained on diverse human video data can enable data-efficient learning of downstream robotic manipulation tasks. Concretely, we pre-train a visual representation using the Ego4D human video dataset using a combination of time-contrastive learning, video-language alignment, and an L1 penalty to encourage sparse and compact representations. The resulting representation, R3M, can be used as a frozen perception module for downstream policy learning. Across a suite of 12 simulated robot manipulation tasks, we find that R3M improves task success by over 20% compared to training from scratch and by over 10% compared to state-of-the-art visual representations like CLIP and MoCo. Furthermore, R3M enables a Franka Emika Panda arm to learn a range of manipulation tasks in a real, cluttered apartment given just 20 demonstrations. Code and pre-trained models are available at https://tinyurl.com/robotr3m.

Abhinav Gupta, Pierre F. J. Lermusiaux · mit

Improving the predictive capability and computational cost of dynamical models is often at the heart of augmenting computational physics with machine learning (ML). However, most learning results are limited in interpretability and generalization over different computational grid resolutions, initial and boundary conditions, domain geometries, and physical or problem-specific parameters. In the present study, we simultaneously address all these challenges by developing the novel and versatile methodology of unified neural partial delay differential equations. We augment existing/low-fidelity dynamical models directly in their partial differential equation (PDE) forms with both Markovian and non-Markovian neural network (NN) closure parameterizations. The melding of the existing models with NNs in the continuous spatiotemporal space followed by numerical discretization automatically allows for the desired generalizability. The Markovian term is designed to enable extraction of its analytical form and thus provides interpretability. The non-Markovian terms allow accounting for inherently missing time delays needed to represent the real world. We obtain adjoint PDEs in the continuous form, thus enabling direct implementation across differentiable and non-differentiable computational physics codes, different ML frameworks, and treatment of nonuniformly-spaced spatiotemporal training data. We demonstrate the new generalized neural closure models (gnCMs) framework using four sets of experiments based on advecting nonlinear waves, shocks, and ocean acidification models. Our learned gnCMs discover missing physics, find leading numerical error terms, discriminate among candidate functional forms in an interpretable fashion, achieve generalization, and compensate for the lack of complexity in simpler models. Finally, we analyze the computational advantages of our new framework.

Simone Parisi, Aravind Rajeswaran, Senthil Purushwalkam et al.

Recent years have seen the emergence of pre-trained representations as a powerful abstraction for AI applications in computer vision, natural language, and speech. However, policy learning for control is still dominated by a tabula-rasa learning paradigm, with visuo-motor policies often trained from scratch using data from deployment environments. In this context, we revisit and study the role of pre-trained visual representations for control, and in particular representations trained on large-scale computer vision datasets. Through extensive empirical evaluation in diverse control domains (Habitat, DeepMind Control, Adroit, Franka Kitchen), we isolate and study the importance of different representation training methods, data augmentations, and feature hierarchies. Overall, we find that pre-trained visual representations can be competitive or even better than ground-truth state representations to train control policies. This is in spite of using only out-of-domain data from standard vision datasets, without any in-domain data from the deployment environments. Source code and more at https://sites.google.com/view/pvr-control.

Yufei Ye, Abhinav Gupta, Shubham Tulsiani

Our work aims to reconstruct hand-held objects given a single RGB image. In contrast to prior works that typically assume known 3D templates and reduce the problem to 3D pose estimation, our work reconstructs generic hand-held object without knowing their 3D templates. Our key insight is that hand articulation is highly predictive of the object shape, and we propose an approach that conditionally reconstructs the object based on the articulation and the visual input. Given an image depicting a hand-held object, we first use off-the-shelf systems to estimate the underlying hand pose and then infer the object shape in a normalized hand-centric coordinate frame. We parameterized the object by signed distance which are inferred by an implicit network which leverages the information from both visual feature and articulation-aware coordinates to process a query point. We perform experiments across three datasets and show that our method consistently outperforms baselines and is able to reconstruct a diverse set of objects. We analyze the benefits and robustness of explicit articulation conditioning and also show that this allows the hand pose estimation to further improve in test-time optimization.

Homanga Bharadhwaj, Jay Vakil, Mohit Sharma et al.

The grand aim of having a single robot that can manipulate arbitrary objects in diverse settings is at odds with the paucity of robotics datasets. Acquiring and growing such datasets is strenuous due to manual efforts, operational costs, and safety challenges. A path toward such an universal agent would require a structured framework capable of wide generalization but trained within a reasonable data budget. In this paper, we develop an efficient system (RoboAgent) for training universal agents capable of multi-task manipulation skills using (a) semantic augmentations that can rapidly multiply existing datasets and (b) action representations that can extract performant policies with small yet diverse multi-modal datasets without overfitting. In addition, reliable task conditioning and an expressive policy architecture enable our agent to exhibit a diverse repertoire of skills in novel situations specified using language commands. Using merely 7500 demonstrations, we are able to train a single agent capable of 12 unique skills, and demonstrate its generalization over 38 tasks spread across common daily activities in diverse kitchen scenes. On average, RoboAgent outperforms prior methods by over 40% in unseen situations while being more sample efficient and being amenable to capability improvements and extensions through fine-tuning. Videos at https://robopen.github.io/

Himangi Mittal, Pedro Morgado, Unnat Jain et al.

We propose a self-supervised algorithm to learn representations from egocentric video data. Recently, significant efforts have been made to capture humans interacting with their own environments as they go about their daily activities. In result, several large egocentric datasets of interaction-rich multi-modal data have emerged. However, learning representations from videos can be challenging. First, given the uncurated nature of long-form continuous videos, learning effective representations require focusing on moments in time when interactions take place. Second, visual representations of daily activities should be sensitive to changes in the state of the environment. However, current successful multi-modal learning frameworks encourage representation invariance over time. To address these challenges, we leverage audio signals to identify moments of likely interactions which are conducive to better learning. We also propose a novel self-supervised objective that learns from audible state changes caused by interactions. We validate these contributions extensively on two large-scale egocentric datasets, EPIC-Kitchens-100 and the recently released Ego4D, and show improvements on several downstream tasks, including action recognition, long-term action anticipation, and object state change classification.

Yufei Ye, Xueting Li, Abhinav Gupta et al.

Recent successes in image synthesis are powered by large-scale diffusion models. However, most methods are currently limited to either text- or image-conditioned generation for synthesizing an entire image, texture transfer or inserting objects into a user-specified region. In contrast, in this work we focus on synthesizing complex interactions (ie, an articulated hand) with a given object. Given an RGB image of an object, we aim to hallucinate plausible images of a human hand interacting with it. We propose a two-step generative approach: a LayoutNet that samples an articulation-agnostic hand-object-interaction layout, and a ContentNet that synthesizes images of a hand grasping the object given the predicted layout. Both are built on top of a large-scale pretrained diffusion model to make use of its latent representation. Compared to baselines, the proposed method is shown to generalize better to novel objects and perform surprisingly well on out-of-distribution in-the-wild scenes of portable-sized objects. The resulting system allows us to predict descriptive affordance information, such as hand articulation and approaching orientation. Project page: https://judyye.github.io/affordiffusion-www

Yufei Ye, Poorvi Hebbar, Abhinav Gupta et al.

We tackle the task of reconstructing hand-object interactions from short video clips. Given an input video, our approach casts 3D inference as a per-video optimization and recovers a neural 3D representation of the object shape, as well as the time-varying motion and hand articulation. While the input video naturally provides some multi-view cues to guide 3D inference, these are insufficient on their own due to occlusions and limited viewpoint variations. To obtain accurate 3D, we augment the multi-view signals with generic data-driven priors to guide reconstruction. Specifically, we learn a diffusion network to model the conditional distribution of (geometric) renderings of objects conditioned on hand configuration and category label, and leverage it as a prior to guide the novel-view renderings of the reconstructed scene. We empirically evaluate our approach on egocentric videos across 6 object categories, and observe significant improvements over prior single-view and multi-view methods. Finally, we demonstrate our system's ability to reconstruct arbitrary clips from YouTube, showing both 1st and 3rd person interactions.

Homanga Bharadhwaj, Abhinav Gupta, Shubham Tulsiani et al.

Can we learn robot manipulation for everyday tasks, only by watching videos of humans doing arbitrary tasks in different unstructured settings? Unlike widely adopted strategies of learning task-specific behaviors or direct imitation of a human video, we develop a a framework for extracting agent-agnostic action representations from human videos, and then map it to the agent's embodiment during deployment. Our framework is based on predicting plausible human hand trajectories given an initial image of a scene. After training this prediction model on a diverse set of human videos from the internet, we deploy the trained model zero-shot for physical robot manipulation tasks, after appropriate transformations to the robot's embodiment. This simple strategy lets us solve coarse manipulation tasks like opening and closing drawers, pushing, and tool use, without access to any in-domain robot manipulation trajectories. Our real-world deployment results establish a strong baseline for action prediction information that can be acquired from diverse arbitrary videos of human activities, and be useful for zero-shot robotic manipulation in unseen scenes.

Justin Wasserman, Karmesh Yadav, Girish Chowdhary et al.

Realistic long-horizon tasks like image-goal navigation involve exploratory and exploitative phases. Assigned with an image of the goal, an embodied agent must explore to discover the goal, i.e., search efficiently using learned priors. Once the goal is discovered, the agent must accurately calibrate the last-mile of navigation to the goal. As with any robust system, switches between exploratory goal discovery and exploitative last-mile navigation enable better recovery from errors. Following these intuitive guide rails, we propose SLING to improve the performance of existing image-goal navigation systems. Entirely complementing prior methods, we focus on last-mile navigation and leverage the underlying geometric structure of the problem with neural descriptors. With simple but effective switches, we can easily connect SLING with heuristic, reinforcement learning, and neural modular policies. On a standardized image-goal navigation benchmark (Hahn et al. 2021), we improve performance across policies, scenes, and episode complexity, raising the state-of-the-art from 45% to 55% success rate. Beyond photorealistic simulation, we conduct real-robot experiments in three physical scenes and find these improvements to transfer well to real environments.

Sudeep Dasari, Mohan Kumar Srirama, Unnat Jain et al.

Visual representation learning hold great promise for robotics, but is severely hampered by the scarcity and homogeneity of robotics datasets. Recent works address this problem by pre-training visual representations on large-scale but out-of-domain data (e.g., videos of egocentric interactions) and then transferring them to target robotics tasks. While the field is heavily focused on developing better pre-training algorithms, we find that dataset choice is just as important to this paradigm's success. After all, the representation can only learn the structures or priors present in the pre-training dataset. To this end, we flip the focus on algorithms, and instead conduct a dataset centric analysis of robotic pre-training. Our findings call into question some common wisdom in the field. We observe that traditional vision datasets (like ImageNet, Kinetics and 100 Days of Hands) are surprisingly competitive options for visuo-motor representation learning, and that the pre-training dataset's image distribution matters more than its size. Finally, we show that common simulation benchmarks are not a reliable proxy for real world performance and that simple regularization strategies can dramatically improve real world policy learning. https://data4robotics.github.io

Abhinav Gupta, Toben H. Mintz, Jesse Thomason

While word embeddings derive meaning from co-occurrence patterns, human language understanding is grounded in sensory and motor experience. We present $\text{SENSE}$ $(\textbf{S}\text{ensorimotor }$ $\textbf{E}\text{mbedding }$ $\textbf{N}\text{orm }$ $\textbf{S}\text{coring }$ $\textbf{E}\text{ngine})$, a learned projection model that predicts Lancaster sensorimotor norms from word lexical embeddings. We also conducted a behavioral study where 281 participants selected which among candidate nonce words evoked specific sensorimotor associations, finding statistically significant correlations between human selection rates and $\text{SENSE}$ ratings across 6 of the 11 modalities. Sublexical analysis of these nonce words selection rates revealed systematic phonosthemic patterns for the interoceptive norm, suggesting a path towards computationally proposing candidate phonosthemes from text data.

Anurag Pratik, Soumith Chintala, Kavya Srinet et al.

In recent years, there have been significant advances in building end-to-end Machine Learning (ML) systems that learn at scale. But most of these systems are: (a) isolated (perception, speech, or language only); (b) trained on static datasets. On the other hand, in the field of robotics, large-scale learning has always been difficult. Supervision is hard to gather and real world physical interactions are expensive. In this work we introduce and open-source droidlet, a modular, heterogeneous agent architecture and platform. It allows us to exploit both large-scale static datasets in perception and language and sophisticated heuristics often used in robotics; and provides tools for interactive annotation. Furthermore, it brings together perception, language and action onto one platform, providing a path towards agents that learn from the richness of real world interactions.

Homanga Bharadhwaj, Abhinav Gupta, Shubham Tulsiani

Motivated by the intuitive understanding humans have about the space of possible interactions, and the ease with which they can generalize this understanding to previously unseen scenes, we develop an approach for learning visual affordances for guiding robot exploration. Given an input image of a scene, we infer a distribution over plausible future states that can be achieved via interactions with it. We use a Transformer-based model to learn a conditional distribution in the latent embedding space of a VQ-VAE and show that these models can be trained using large-scale and diverse passive data, and that the learned models exhibit compositional generalization to diverse objects beyond the training distribution. We show how the trained affordance model can be used for guiding exploration by acting as a goal-sampling distribution, during visual goal-conditioned policy learning in robotic manipulation.

Pushkal Katara, Y V S Harish, Harit Pandya et al.

The simplicity of the visual servoing approach makes it an attractive option for tasks dealing with vision-based control of robots in many real-world applications. However, attaining precise alignment for unseen environments pose a challenge to existing visual servoing approaches. While classical approaches assume a perfect world, the recent data-driven approaches face issues when generalizing to novel environments. In this paper, we aim to combine the best of both worlds. We present a deep model predictive visual servoing framework that can achieve precise alignment with optimal trajectories and can generalize to novel environments. Our framework consists of a deep network for optical flow predictions, which are used along with a predictive model to forecast future optical flow. For generating an optimal set of velocities we present a control network that can be trained on the fly without any supervision. Through extensive simulations on photo-realistic indoor settings of the popular Habitat framework, we show significant performance gain due to the proposed formulation vis-a-vis recent state-of-the-art methods. Specifically, we show a faster convergence and an improved performance in trajectory length over recent approaches.

Shubham Tulsiani, Abhinav Gupta

We propose a generative model that can infer a distribution for the underlying spatial signal conditioned on sparse samples e.g. plausible images given a few observed pixels. In contrast to sequential autoregressive generative models, our model allows conditioning on arbitrary samples and can answer distributional queries for any location. We empirically validate our approach across three image datasets and show that we learn to generate diverse and meaningful samples, with the distribution variance reducing given more observed pixels. We also show that our approach is applicable beyond images and can allow generating other types of spatial outputs e.g. polynomials, 3D shapes, and videos.

Kaichun Mo, Leonidas Guibas, Mustafa Mukadam et al.

One of the fundamental goals of visual perception is to allow agents to meaningfully interact with their environment. In this paper, we take a step towards that long-term goal -- we extract highly localized actionable information related to elementary actions such as pushing or pulling for articulated objects with movable parts. For example, given a drawer, our network predicts that applying a pulling force on the handle opens the drawer. We propose, discuss, and evaluate novel network architectures that given image and depth data, predict the set of actions possible at each pixel, and the regions over articulated parts that are likely to move under the force. We propose a learning-from-interaction framework with an online data sampling strategy that allows us to train the network in simulation (SAPIEN) and generalizes across categories. Check the website for code and data release: https://cs.stanford.edu/~kaichun/where2act/

Kenneth Marino, Xinlei Chen, Devi Parikh et al.

One of the most challenging question types in VQA is when answering the question requires outside knowledge not present in the image. In this work we study open-domain knowledge, the setting when the knowledge required to answer a question is not given/annotated, neither at training nor test time. We tap into two types of knowledge representations and reasoning. First, implicit knowledge which can be learned effectively from unsupervised language pre-training and supervised training data with transformer-based models. Second, explicit, symbolic knowledge encoded in knowledge bases. Our approach combines both - exploiting the powerful implicit reasoning of transformer models for answer prediction, and integrating symbolic representations from a knowledge graph, while never losing their explicit semantics to an implicit embedding. We combine diverse sources of knowledge to cover the wide variety of knowledge needed to solve knowledge-based questions. We show our approach, KRISP (Knowledge Reasoning with Implicit and Symbolic rePresentations), significantly outperforms state-of-the-art on OK-VQA, the largest available dataset for open-domain knowledge-based VQA. We show with extensive ablations that while our model successfully exploits implicit knowledge reasoning, the symbolic answer module which explicitly connects the knowledge graph to the answer vocabulary is critical to the performance of our method and generalizes to rare answers.

Shubham Tulsiani, Nilesh Kulkarni, Abhinav Gupta

We present an approach to infer the 3D shape, texture, and camera pose for an object from a single RGB image, using only category-level image collections with foreground masks as supervision. We represent the shape as an image-conditioned implicit function that transforms the surface of a sphere to that of the predicted mesh, while additionally predicting the corresponding texture. To derive supervisory signal for learning, we enforce that: a) our predictions when rendered should explain the available image evidence, and b) the inferred 3D structure should be geometrically consistent with learned pixel to surface mappings. We empirically show that our approach improves over prior work that leverages similar supervision, and in fact performs competitively to methods that use stronger supervision. Finally, as our method enables learning with limited supervision, we qualitatively demonstrate its applicability over a set of about 30 object categories.

Devendra Singh Chaplot, Ruslan Salakhutdinov, Abhinav Gupta et al.

This paper studies the problem of image-goal navigation which involves navigating to the location indicated by a goal image in a novel previously unseen environment. To tackle this problem, we design topological representations for space that effectively leverage semantics and afford approximate geometric reasoning. At the heart of our representations are nodes with associated semantic features, that are interconnected using coarse geometric information. We describe supervised learning-based algorithms that can build, maintain and use such representations under noisy actuation. Experimental study in visually and physically realistic simulation suggests that our method builds effective representations that capture structural regularities and efficiently solve long-horizon navigation problems. We observe a relative improvement of more than 50% over existing methods that study this task.

Nilesh Kulkarni, Abhinav Gupta, David F. Fouhey et al.

We tackle the tasks of: 1) predicting a Canonical Surface Mapping (CSM) that indicates the mapping from 2D pixels to corresponding points on a canonical template shape, and 2) inferring the articulation and pose of the template corresponding to the input image. While previous approaches rely on keypoint supervision for learning, we present an approach that can learn without such annotations. Our key insight is that these tasks are geometrically related, and we can obtain supervisory signal via enforcing consistency among the predictions. We present results across a diverse set of animal object categories, showing that our method can learn articulation and CSM prediction from image collections using only foreground mask labels for training. We empirically show that allowing articulation helps learn more accurate CSM prediction, and that enforcing the consistency with predicted CSM is similarly critical for learning meaningful articulation.

Kiana Ehsani, Shubham Tulsiani, Saurabh Gupta et al.

When we humans look at a video of human-object interaction, we can not only infer what is happening but we can even extract actionable information and imitate those interactions. On the other hand, current recognition or geometric approaches lack the physicality of action representation. In this paper, we take a step towards a more physical understanding of actions. We address the problem of inferring contact points and the physical forces from videos of humans interacting with objects. One of the main challenges in tackling this problem is obtaining ground-truth labels for forces. We sidestep this problem by instead using a physics simulator for supervision. Specifically, we use a simulator to predict effects and enforce that estimated forces must lead to the same effect as depicted in the video. Our quantitative and qualitative results show that (a) we can predict meaningful forces from videos whose effects lead to accurate imitation of the motions observed, (b) by jointly optimizing for contact point and force prediction, we can improve the performance on both tasks in comparison to independent training, and (c) we can learn a representation from this model that generalizes to novel objects using few shot examples.

Pratyusha Sharma, Deepak Pathak, Abhinav Gupta

We study a generalized setup for learning from demonstration to build an agent that can manipulate novel objects in unseen scenarios by looking at only a single video of human demonstration from a third-person perspective. To accomplish this goal, our agent should not only learn to understand the intent of the demonstrated third-person video in its context but also perform the intended task in its environment configuration. Our central insight is to enforce this structure explicitly during learning by decoupling what to achieve (intended task) from how to perform it (controller). We propose a hierarchical setup where a high-level module learns to generate a series of first-person sub-goals conditioned on the third-person video demonstration, and a low-level controller predicts the actions to achieve those sub-goals. Our agent acts from raw image observations without any access to the full state information. We show results on a real robotic platform using Baxter for the manipulation tasks of pouring and placing objects in a box. Project video and code are at https://pathak22.github.io/hierarchical-imitation/

Yufei Ye, Maneesh Singh, Abhinav Gupta et al.

We present an approach for pixel-level future prediction given an input image of a scene. We observe that a scene is comprised of distinct entities that undergo motion and present an approach that operationalizes this insight. We implicitly predict future states of independent entities while reasoning about their interactions, and compose future video frames using these predicted states. We overcome the inherent multi-modality of the task using a global trajectory-level latent random variable, and show that this allows us to sample diverse and plausible futures. We empirically validate our approach against alternate representations and ways of incorporating multi-modality. We examine two datasets, one comprising of stacked objects that may fall, and the other containing videos of humans performing activities in a gym, and show that our approach allows realistic stochastic video prediction across these diverse settings. See https://judyye.github.io/CVP/ for video predictions.

Nilesh Kulkarni, Abhinav Gupta, Shubham Tulsiani

We explore the task of Canonical Surface Mapping (CSM). Specifically, given an image, we learn to map pixels on the object to their corresponding locations on an abstract 3D model of the category. But how do we learn such a mapping? A supervised approach would require extensive manual labeling which is not scalable beyond a few hand-picked categories. Our key insight is that the CSM task (pixel to 3D), when combined with 3D projection (3D to pixel), completes a cycle. Hence, we can exploit a geometric cycle consistency loss, thereby allowing us to forgo the dense manual supervision. Our approach allows us to train a CSM model for a diverse set of classes, without sparse or dense keypoint annotation, by leveraging only foreground mask labels for training. We show that our predictions also allow us to infer dense correspondence between two images, and compare the performance of our approach against several methods that predict correspondence by leveraging varying amount of supervision.

Xiaolong Wang, Abhinav Gupta

How do humans recognize the action "opening a book" ? We argue that there are two important cues: modeling temporal shape dynamics and modeling functional relationships between humans and objects. In this paper, we propose to represent videos as space-time region graphs which capture these two important cues. Our graph nodes are defined by the object region proposals from different frames in a long range video. These nodes are connected by two types of relations: (i) similarity relations capturing the long range dependencies between correlated objects and (ii) spatial-temporal relations capturing the interactions between nearby objects. We perform reasoning on this graph representation via Graph Convolutional Networks. We achieve state-of-the-art results on both Charades and Something-Something datasets. Especially for Charades, we obtain a huge 4.4% gain when our model is applied in complex environments.

Gunnar A. Sigurdsson, Abhinav Gupta, Cordelia Schmid et al.

In Actor and Observer we introduced a dataset linking the first and third-person video understanding domains, the Charades-Ego Dataset. In this paper we describe the egocentric aspect of the dataset and present annotations for Charades-Ego with 68,536 activity instances in 68.8 hours of first and third-person video, making it one of the largest and most diverse egocentric datasets available. Charades-Ego furthermore shares activity classes, scripts, and methodology with the Charades dataset, that consist of additional 82.3 hours of third-person video with 66,500 activity instances. Charades-Ego has temporal annotations and textual descriptions, making it suitable for egocentric video classification, localization, captioning, and new tasks utilizing the cross-modal nature of the data.

Xinlei Chen, Li-Jia Li, Li Fei-Fei et al.

We present a novel framework for iterative visual reasoning. Our framework goes beyond current recognition systems that lack the capability to reason beyond stack of convolutions. The framework consists of two core modules: a local module that uses spatial memory to store previous beliefs with parallel updates; and a global graph-reasoning module. Our graph module has three components: a) a knowledge graph where we represent classes as nodes and build edges to encode different types of semantic relationships between them; b) a region graph of the current image where regions in the image are nodes and spatial relationships between these regions are edges; c) an assignment graph that assigns regions to classes. Both the local module and the global module roll-out iteratively and cross-feed predictions to each other to refine estimates. The final predictions are made by combining the best of both modules with an attention mechanism. We show strong performance over plain ConvNets, \eg achieving an $8.4\%$ absolute improvement on ADE measured by per-class average precision. Analysis also shows that the framework is resilient to missing regions for reasoning.

Chen Sun, Abhinav Shrivastava, Saurabh Singh et al.

The success of deep learning in vision can be attributed to: (a) models with high capacity; (b) increased computational power; and (c) availability of large-scale labeled data. Since 2012, there have been significant advances in representation capabilities of the models and computational capabilities of GPUs. But the size of the biggest dataset has surprisingly remained constant. What will happen if we increase the dataset size by 10x or 100x? This paper takes a step towards clearing the clouds of mystery surrounding the relationship between `enormous data' and visual deep learning. By exploiting the JFT-300M dataset which has more than 375M noisy labels for 300M images, we investigate how the performance of current vision tasks would change if this data was used for representation learning. Our paper delivers some surprising (and some expected) findings. First, we find that the performance on vision tasks increases logarithmically based on volume of training data size. Second, we show that representation learning (or pre-training) still holds a lot of promise. One can improve performance on many vision tasks by just training a better base model. Finally, as expected, we present new state-of-the-art results for different vision tasks including image classification, object detection, semantic segmentation and human pose estimation. Our sincere hope is that this inspires vision community to not undervalue the data and develop collective efforts in building larger datasets.

Xinlei Chen, Abhinav Gupta

Modeling instance-level context and object-object relationships is extremely challenging. It requires reasoning about bounding boxes of different classes, locations \etc. Above all, instance-level spatial reasoning inherently requires modeling conditional distributions on previous detections. Unfortunately, our current object detection systems do not have any {\bf memory} to remember what to condition on! The state-of-the-art object detectors still detect all object in parallel followed by non-maximal suppression (NMS). While memory has been used for tasks such as captioning, they mostly use image-level memory cells without capturing the spatial layout. On the other hand, modeling object-object relationships requires {\bf spatial} reasoning -- not only do we need a memory to store the spatial layout, but also a effective reasoning module to extract spatial patterns. This paper presents a conceptually simple yet powerful solution -- Spatial Memory Network (SMN), to model the instance-level context efficiently and effectively. Our spatial memory essentially assembles object instances back into a pseudo "image" representation that is easy to be fed into another ConvNet for object-object context reasoning. This leads to a new sequential reasoning architecture where image and memory are processed in parallel to obtain detections which update the memory again. We show our SMN direction is promising as it provides 2.2\% improvement over baseline Faster RCNN on the COCO dataset so far.

Siddha Ganju, Olga Russakovsky, Abhinav Gupta

Collecting fully annotated image datasets is challenging and expensive. Many types of weak supervision have been explored: weak manual annotations, web search results, temporal continuity, ambient sound and others. We focus on one particular unexplored mode: visual questions that are asked about images. The key observation that inspires our work is that the question itself provides useful information about the image (even without the answer being available). For instance, the question "what is the breed of the dog?" informs the AI that the animal in the scene is a dog and that there is only one dog present. We make three contributions: (1) providing an extensive qualitative and quantitative analysis of the information contained in human visual questions, (2) proposing two simple but surprisingly effective modifications to the standard visual question answering models that allow them to make use of weak supervision in the form of unanswered questions associated with images and (3) demonstrating that a simple data augmentation strategy inspired by our insights results in a 7.1% improvement on the standard VQA benchmark.

Xiaolong Wang, Abhinav Shrivastava, Abhinav Gupta

How do we learn an object detector that is invariant to occlusions and deformations? Our current solution is to use a data-driven strategy -- collect large-scale datasets which have object instances under different conditions. The hope is that the final classifier can use these examples to learn invariances. But is it really possible to see all the occlusions in a dataset? We argue that like categories, occlusions and object deformations also follow a long-tail. Some occlusions and deformations are so rare that they hardly happen; yet we want to learn a model invariant to such occurrences. In this paper, we propose an alternative solution. We propose to learn an adversarial network that generates examples with occlusions and deformations. The goal of the adversary is to generate examples that are difficult for the object detector to classify. In our framework both the original detector and adversary are learned in a joint manner. Our experimental results indicate a 2.3% mAP boost on VOC07 and a 2.6% mAP boost on VOC2012 object detection challenge compared to the Fast-RCNN pipeline. We also release the code for this paper.

Aayush Bansal, Xinlei Chen, Bryan Russell et al.

We explore design principles for general pixel-level prediction problems, from low-level edge detection to mid-level surface normal estimation to high-level semantic segmentation. Convolutional predictors, such as the fully-convolutional network (FCN), have achieved remarkable success by exploiting the spatial redundancy of neighboring pixels through convolutional processing. Though computationally efficient, we point out that such approaches are not statistically efficient during learning precisely because spatial redundancy limits the information learned from neighboring pixels. We demonstrate that stratified sampling of pixels allows one to (1) add diversity during batch updates, speeding up learning; (2) explore complex nonlinear predictors, improving accuracy; and (3) efficiently train state-of-the-art models tabula rasa (i.e., "from scratch") for diverse pixel-labeling tasks. Our single architecture produces state-of-the-art results for semantic segmentation on PASCAL-Context dataset, surface normal estimation on NYUDv2 depth dataset, and edge detection on BSDS.

Xinlei Chen, Abhinav Gupta

We adapted the join-training scheme of Faster RCNN framework from Caffe to TensorFlow as a baseline implementation for object detection. Our code is made publicly available. This report documents the simplifications made to the original pipeline, with justifications from ablation analysis on both PASCAL VOC 2007 and COCO 2014. We further investigated the role of non-maximal suppression (NMS) in selecting regions-of-interest (RoIs) for region classification, and found that a biased sampling toward small regions helps performance and can achieve on-par mAP to NMS-based sampling when converged sufficiently.

Abhinav Shrivastava, Rahul Sukthankar, Jitendra Malik et al.

In recent years, we have seen tremendous progress in the field of object detection. Most of the recent improvements have been achieved by targeting deeper feedforward networks. However, many hard object categories such as bottle, remote, etc. require representation of fine details and not just coarse, semantic representations. But most of these fine details are lost in the early convolutional layers. What we need is a way to incorporate finer details from lower layers into the detection architecture. Skip connections have been proposed to combine high-level and low-level features, but we argue that selecting the right features from low-level requires top-down contextual information. Inspired by the human visual pathway, in this paper we propose top-down modulations as a way to incorporate fine details into the detection framework. Our approach supplements the standard bottom-up, feedforward ConvNet with a top-down modulation (TDM) network, connected using lateral connections. These connections are responsible for the modulation of lower layer filters, and the top-down network handles the selection and integration of contextual information and low-level features. The proposed TDM architecture provides a significant boost on the COCO testdev benchmark, achieving 28.6 AP for VGG16, 35.2 AP for ResNet101, and 37.3 for InceptionResNetv2 network, without any bells and whistles (e.g., multi-scale, iterative box refinement, etc.).

Gunnar A. Sigurdsson, Santosh Divvala, Ali Farhadi et al.

Actions are more than just movements and trajectories: we cook to eat and we hold a cup to drink from it. A thorough understanding of videos requires going beyond appearance modeling and necessitates reasoning about the sequence of activities, as well as the higher-level constructs such as intentions. But how do we model and reason about these? We propose a fully-connected temporal CRF model for reasoning over various aspects of activities that includes objects, actions, and intentions, where the potentials are predicted by a deep network. End-to-end training of such structured models is a challenging endeavor: For inference and learning we need to construct mini-batches consisting of whole videos, leading to mini-batches with only a few videos. This causes high-correlation between data points leading to breakdown of the backprop algorithm. To address this challenge, we present an asynchronous variational inference method that allows efficient end-to-end training. Our method achieves a classification mAP of 22.4% on the Charades benchmark, outperforming the state-of-the-art (17.2% mAP), and offers equal gains on the task of temporal localization.

Aayush Bansal, Xinlei Chen, Bryan Russell et al.

We explore architectures for general pixel-level prediction problems, from low-level edge detection to mid-level surface normal estimation to high-level semantic segmentation. Convolutional predictors, such as the fully-convolutional network (FCN), have achieved remarkable success by exploiting the spatial redundancy of neighboring pixels through convolutional processing. Though computationally efficient, we point out that such approaches are not statistically efficient during learning precisely because spatial redundancy limits the information learned from neighboring pixels. We demonstrate that (1) stratified sampling allows us to add diversity during batch updates and (2) sampled multi-scale features allow us to explore more nonlinear predictors (multiple fully-connected layers followed by ReLU) that improve overall accuracy. Finally, our objective is to show how a architecture can get performance better than (or comparable to) the architectures designed for a particular task. Interestingly, our single architecture produces state-of-the-art results for semantic segmentation on PASCAL-Context, surface normal estimation on NYUDv2 dataset, and edge detection on BSDS without contextual post-processing.

Senthil Purushwalkam, Abhinav Gupta

Human actions are comprised of a sequence of poses. This makes videos of humans a rich and dense source of human poses. We propose an unsupervised method to learn pose features from videos that exploits a signal which is complementary to appearance and can be used as supervision: motion. The key idea is that humans go through poses in a predictable manner while performing actions. Hence, given two poses, it should be possible to model the motion that caused the change between them. We represent each of the poses as a feature in a CNN (Appearance ConvNet) and generate a motion encoding from optical flow maps using a separate CNN (Motion ConvNet). The data for this task is automatically generated allowing us to train without human supervision. We demonstrate the strength of the learned representation by finetuning the trained model for Pose Estimation on the FLIC dataset, for static image action recognition on PASCAL and for action recognition in videos on UCF101 and HMDB51.

Gunnar A. Sigurdsson, Olga Russakovsky, Ali Farhadi et al.

Large-scale annotated datasets allow AI systems to learn from and build upon the knowledge of the crowd. Many crowdsourcing techniques have been developed for collecting image annotations. These techniques often implicitly rely on the fact that a new input image takes a negligible amount of time to perceive. In contrast, we investigate and determine the most cost-effective way of obtaining high-quality multi-label annotations for temporal data such as videos. Watching even a short 30-second video clip requires a significant time investment from a crowd worker; thus, requesting multiple annotations following a single viewing is an important cost-saving strategy. But how many questions should we ask per video? We conclude that the optimal strategy is to ask as many questions as possible in a HIT (up to 52 binary questions after watching a 30-second video clip in our experiments). We demonstrate that while workers may not correctly answer all questions, the cost-benefit analysis nevertheless favors consensus from multiple such cheap-yet-imperfect iterations over more complex alternatives. When compared with a one-question-per-video baseline, our method is able to achieve a 10% improvement in recall 76.7% ours versus 66.7% baseline) at comparable precision (83.8% ours versus 83.0% baseline) in about half the annotation time (3.8 minutes ours compared to 7.1 minutes baseline). We demonstrate the effectiveness of our method by collecting multi-label annotations of 157 human activities on 1,815 videos.

Xiaolong Wang, Abhinav Gupta

Current generative frameworks use end-to-end learning and generate images by sampling from uniform noise distribution. However, these approaches ignore the most basic principle of image formation: images are product of: (a) Structure: the underlying 3D model; (b) Style: the texture mapped onto structure. In this paper, we factorize the image generation process and propose Style and Structure Generative Adversarial Network (S^2-GAN). Our S^2-GAN has two components: the Structure-GAN generates a surface normal map; the Style-GAN takes the surface normal map as input and generates the 2D image. Apart from a real vs. generated loss function, we use an additional loss with computed surface normals from generated images. The two GANs are first trained independently, and then merged together via joint learning. We show our S^2-GAN model is interpretable, generates more realistic images and can be used to learn unsupervised RGBD representations.

David F. Fouhey, Xiaolong Wang, Abhinav Gupta

The field of functional recognition or affordance estimation from images has seen a revival in recent years. As originally proposed by Gibson, the affordances of a scene were directly perceived from the ambient light: in other words, functional properties like sittable were estimated directly from incoming pixels. Recent work, however, has taken a mediated approach in which affordances are derived by first estimating semantics or geometry and then reasoning about the affordances. In a tribute to Gibson, this paper explores his theory of affordances as originally proposed. We propose two approaches for direct perception of affordances and show that they obtain good results and can out-perform mediated approaches. We hope this paper can rekindle discussion around direct perception and its implications in the long term.

Aayush Bansal, Abhinav Shrivastava, Carl Doersch et al.

Building on the success of recent discriminative mid-level elements, we propose a surprisingly simple approach for object detection which performs comparable to the current state-of-the-art approaches on PASCAL VOC comp-3 detection challenge (no external data). Through extensive experiments and ablation analysis, we show how our approach effectively improves upon the HOG-based pipelines by adding an intermediate mid-level representation for the task of object detection. This representation is easily interpretable and allows us to visualize what our object detector "sees". We also discuss the insights our approach shares with CNN-based methods, such as sharing representation between categories helps.

Naiyan Wang, Siyi Li, Abhinav Gupta et al.

Convolutional neural network (CNN) models have demonstrated great success in various computer vision tasks including image classification and object detection. However, some equally important tasks such as visual tracking remain relatively unexplored. We believe that a major hurdle that hinders the application of CNN to visual tracking is the lack of properly labeled training data. While existing applications that liberate the power of CNN often need an enormous amount of training data in the order of millions, visual tracking applications typically have only one labeled example in the first frame of each video. We address this research issue here by pre-training a CNN offline and then transferring the rich feature hierarchies learned to online tracking. The CNN is also fine-tuned during online tracking to adapt to the appearance of the tracked target specified in the first video frame. To fit the characteristics of object tracking, we first pre-train the CNN to recognize what is an object, and then propose to generate a probability map instead of producing a simple class label. Using two challenging open benchmarks for performance evaluation, our proposed tracker has demonstrated substantial improvement over other state-of-the-art trackers.