Forward Prediction for Physical Reasoning

Physical reasoning requires forward prediction: the ability to forecast what will happen next given some initial world state. We study the performance of state-of-the-art forward-prediction models in complex physical-reasoning tasks. We do so by incorporating models that operate on object or pixel-based representations of the world, into simple physical-reasoning agents. We find that forward-prediction models improve the performance of physical-reasoning agents, particularly on complex tasks that involve many objects. However, we also find that these improvements are contingent on the training tasks being similar to the test tasks, and that generalization to different tasks is more challenging. Surprisingly, we observe that forward predictors with better pixel accuracy do not necessarily lead to better physical-reasoning performance. Nevertheless, our best models set a new state-of-the-art on the PHYRE benchmark for physical reasoning.

Video Understanding as Machine Translation

With the advent of large-scale multimodal video datasets, especially sequences with audio or transcribed speech, there has been a growing interest in self-supervised learning of video representations. Most prior work formulates the objective as a contrastive metric learning problem between the modalities. To enable effective learning, however, these strategies require a careful selection of positive and negative samples often combined with hand-designed curriculum policies. In this work we remove the need for negative sampling by taking a generative modeling approach that poses the objective as a translation problem between modalities. Such a formulation allows us to tackle a wide variety of downstream video understanding tasks by means of a single unified framework, without the need for large batches of negative samples common in contrastive metric learning. We experiment with the large-scale HowTo100M dataset for training, and report performance gains over the state-of-the-art on several downstream tasks including video classification (EPIC-Kitchens), question answering (TVQA), captioning (TVC, YouCook2, and MSR-VTT), and text-based clip retrieval (YouCook2 and MSR-VTT).

Are we asking the right questions in MovieQA?

Joint vision and language tasks like visual question answering are fascinating because they explore high-level understanding, but at the same time, can be more prone to language biases. In this paper, we explore the biases in the MovieQA dataset and propose a strikingly simple model which can exploit them. We find that using the right word embedding is of utmost importance. By using an appropriately trained word embedding, about half the Question-Answers (QAs) can be answered by looking at the questions and answers alone, completely ignoring narrative context from video clips, subtitles, and movie scripts. Compared to the best published papers on the leaderboard, our simple question + answer only model improves accuracy by 5% for video + subtitle category, 5% for subtitle, 15% for DVS and 6% higher for scripts.

CATER: A diagnostic dataset for Compositional Actions and TEmporal Reasoning

Computer vision has undergone a dramatic revolution in performance, driven in large part through deep features trained on large-scale supervised datasets. However, much of these improvements have focused on static image analysis; video understanding has seen rather modest improvements. Even though new datasets and spatiotemporal models have been proposed, simple frame-by-frame classification methods often still remain competitive. We posit that current video datasets are plagued with implicit biases over scene and object structure that can dwarf variations in temporal structure. In this work, we build a video dataset with fully observable and controllable object and scene bias, and which truly requires spatiotemporal understanding in order to be solved. Our dataset, named CATER, is rendered synthetically using a library of standard 3D objects, and tests the ability to recognize compositions of object movements that require long-term reasoning. In addition to being a challenging dataset, CATER also provides a plethora of diagnostic tools to analyze modern spatiotemporal video architectures by being completely observable and controllable. Using CATER, we provide insights into some of the most recent state of the art deep video architectures.

MetaPix: Few-Shot Video Retargeting

We address the task of unsupervised retargeting of human actions from one video to another. We consider the challenging setting where only a few frames of the target is available. The core of our approach is a conditional generative model that can transcode input skeletal poses (automatically extracted with an off-the-shelf pose estimator) to output target frames. However, it is challenging to build a universal transcoder because humans can appear wildly different due to clothing and background scene geometry. Instead, we learn to adapt - or personalize - a universal generator to the particular human and background in the target. To do so, we make use of meta-learning to discover effective strategies for on-the-fly personalization. One significant benefit of meta-learning is that the personalized transcoder naturally enforces temporal coherence across its generated frames; all frames contain consistent clothing and background geometry of the target. We experiment on in-the-wild internet videos and images and show our approach improves over widely-used baselines for the task.

DistInit: Learning Video Representations without a Single Labeled Video

Video recognition models have progressed significantly over the past few years, evolving from shallow classifiers trained on hand-crafted features to deep spatiotemporal networks. However, labeled video data required to train such models has not been able to keep up with the ever increasing depth and sophistication of these networks. In this work we propose an alternative approach to learning video representations that requires no semantically labeled videos, and instead leverages the years of effort in collecting and labeling large and clean still-image datasets. We do so by using state-of-the-art models pre-trained on image datasets as "teachers" to train video models in a distillation framework. We demonstrate that our method learns truly spatiotemporal features, despite being trained only using supervision from still-image networks. Moreover, it learns good representations across different input modalities, using completely uncurated raw video data sources and with different 2D teacher models. Our method obtains strong transfer performance, outperforming standard techniques for bootstrapping video architectures from image-based models and obtains competitive performance with state-of-the-art approaches for video action recognition.

Video Action Transformer Network

We introduce the Action Transformer model for recognizing and localizing human actions in video clips. We repurpose a Transformer-style architecture to aggregate features from the spatiotemporal context around the person whose actions we are trying to classify. We show that by using high-resolution, person-specific, class-agnostic queries, the model spontaneously learns to track individual people and to pick up on semantic context from the actions of others. Additionally its attention mechanism learns to emphasize hands and faces, which are often crucial to discriminate an action - all without explicit supervision other than boxes and class labels. We train and test our Action Transformer network on the Atomic Visual Actions (AVA) dataset, outperforming the state-of-the-art by a significant margin - more than 7.5% absolute (40% relative) improvement, using only raw RGB frames as input.

A Better Baseline for AVA

We introduce a simple baseline for action localization on the AVA dataset. The model builds upon the Faster R-CNN bounding box detection framework, adapted to operate on pure spatiotemporal features - in our case produced exclusively by an I3D model pretrained on Kinetics. This model obtains 21.9% average AP on the validation set of AVA v2.1, up from 14.5% for the best RGB spatiotemporal model used in the original AVA paper (which was pretrained on Kinetics and ImageNet), and up from 11.3 of the publicly available baseline using a ResNet101 image feature extractor, that was pretrained on ImageNet. Our final model obtains 22.8%/21.9% mAP on the val/test sets and outperforms all submissions to the AVA challenge at CVPR 2018.

Detect-and-Track: Efficient Pose Estimation in Videos

This paper addresses the problem of estimating and tracking human body keypoints in complex, multi-person video. We propose an extremely lightweight yet highly effective approach that builds upon the latest advancements in human detection and video understanding. Our method operates in two-stages: keypoint estimation in frames or short clips, followed by lightweight tracking to generate keypoint predictions linked over the entire video. For frame-level pose estimation we experiment with Mask R-CNN, as well as our own proposed 3D extension of this model, which leverages temporal information over small clips to generate more robust frame predictions. We conduct extensive ablative experiments on the newly released multi-person video pose estimation benchmark, PoseTrack, to validate various design choices of our model. Our approach achieves an accuracy of 55.2% on the validation and 51.8% on the test set using the Multi-Object Tracking Accuracy (MOTA) metric, and achieves state of the art performance on the ICCV 2017 PoseTrack keypoint tracking challenge.

Binge Watching: Scaling Affordance Learning from Sitcoms

In recent years, there has been a renewed interest in jointly modeling perception and action. At the core of this investigation is the idea of modeling affordances(Affordances are opportunities of interaction in the scene. In other words, it represents what actions can the object be used for). However, when it comes to predicting affordances, even the state of the art approaches still do not use any ConvNets. Why is that? Unlike semantic or 3D tasks, there still does not exist any large-scale dataset for affordances. In this paper, we tackle the challenge of creating one of the biggest dataset for learning affordances. We use seven sitcoms to extract a diverse set of scenes and how actors interact with different objects in the scenes. Our dataset consists of more than 10K scenes and 28K ways humans can interact with these 10K images. We also propose a two-step approach to predict affordances in a new scene. In the first step, given a location in the scene we classify which of the 30 pose classes is the likely affordance pose. Given the pose class and the scene, we then use a Variational Autoencoder (VAE) to extract the scale and deformation of the pose. The VAE allows us to sample the distribution of possible poses at test time. Finally, we show the importance of large-scale data in learning a generalizable and robust model of affordances.