Carl Doersch

Klaus Greff, Francois Belletti, Lucas Beyer et al. · deepmind, mila

Data is the driving force of machine learning, with the amount and quality of training data often being more important for the performance of a system than architecture and training details. But collecting, processing and annotating real data at scale is difficult, expensive, and frequently raises additional privacy, fairness and legal concerns. Synthetic data is a powerful tool with the potential to address these shortcomings: 1) it is cheap 2) supports rich ground-truth annotations 3) offers full control over data and 4) can circumvent or mitigate problems regarding bias, privacy and licensing. Unfortunately, software tools for effective data generation are less mature than those for architecture design and training, which leads to fragmented generation efforts. To address these problems we introduce Kubric, an open-source Python framework that interfaces with PyBullet and Blender to generate photo-realistic scenes, with rich annotations, and seamlessly scales to large jobs distributed over thousands of machines, and generating TBs of data. We demonstrate the effectiveness of Kubric by presenting a series of 13 different generated datasets for tasks ranging from studying 3D NeRF models to optical flow estimation. We release Kubric, the used assets, all of the generation code, as well as the rendered datasets for reuse and modification.

Mel Vecerik, Carl Doersch, Yi Yang et al. · deepmind

For robots to be useful outside labs and specialized factories we need a way to teach them new useful behaviors quickly. Current approaches lack either the generality to onboard new tasks without task-specific engineering, or else lack the data-efficiency to do so in an amount of time that enables practical use. In this work we explore dense tracking as a representational vehicle to allow faster and more general learning from demonstration. Our approach utilizes Track-Any-Point (TAP) models to isolate the relevant motion in a demonstration, and parameterize a low-level controller to reproduce this motion across changes in the scene configuration. We show this results in robust robot policies that can solve complex object-arrangement tasks such as shape-matching, stacking, and even full path-following tasks such as applying glue and sticking objects together, all from demonstrations that can be collected in minutes.

Carl Doersch, Yi Yang, Mel Vecerik et al.

We present a novel model for Tracking Any Point (TAP) that effectively tracks any queried point on any physical surface throughout a video sequence. Our approach employs two stages: (1) a matching stage, which independently locates a suitable candidate point match for the query point on every other frame, and (2) a refinement stage, which updates both the trajectory and query features based on local correlations. The resulting model surpasses all baseline methods by a significant margin on the TAP-Vid benchmark, as demonstrated by an approximate 20% absolute average Jaccard (AJ) improvement on DAVIS. Our model facilitates fast inference on long and high-resolution video sequences. On a modern GPU, our implementation has the capacity to track points faster than real-time, and can be flexibly extended to higher-resolution videos. Given the high-quality trajectories extracted from a large dataset, we demonstrate a proof-of-concept diffusion model which generates trajectories from static images, enabling plausible animations. Visualizations, source code, and pretrained models can be found on our project webpage.

Carl Doersch, Ankush Gupta, Larisa Markeeva et al.

Generic motion understanding from video involves not only tracking objects, but also perceiving how their surfaces deform and move. This information is useful to make inferences about 3D shape, physical properties and object interactions. While the problem of tracking arbitrary physical points on surfaces over longer video clips has received some attention, no dataset or benchmark for evaluation existed, until now. In this paper, we first formalize the problem, naming it tracking any point (TAP). We introduce a companion benchmark, TAP-Vid, which is composed of both real-world videos with accurate human annotations of point tracks, and synthetic videos with perfect ground-truth point tracks. Central to the construction of our benchmark is a novel semi-automatic crowdsourced pipeline which uses optical flow estimates to compensate for easier, short-term motion like camera shake, allowing annotators to focus on harder sections of video. We validate our pipeline on synthetic data and propose a simple end-to-end point tracking model TAP-Net, showing that it outperforms all prior methods on our benchmark when trained on synthetic data.

Homanga Bharadhwaj, Debidatta Dwibedi, Abhinav Gupta et al.

How can robot manipulation policies generalize to novel tasks involving unseen object types and new motions? In this paper, we provide a solution in terms of predicting motion information from web data through human video generation and conditioning a robot policy on the generated video. Instead of attempting to scale robot data collection which is expensive, we show how we can leverage video generation models trained on easily available web data, for enabling generalization. Our approach Gen2Act casts language-conditioned manipulation as zero-shot human video generation followed by execution with a single policy conditioned on the generated video. To train the policy, we use an order of magnitude less robot interaction data compared to what the video prediction model was trained on. Gen2Act doesn't require fine-tuning the video model at all and we directly use a pre-trained model for generating human videos. Our results on diverse real-world scenarios show how Gen2Act enables manipulating unseen object types and performing novel motions for tasks not present in the robot data. Videos are at https://homangab.github.io/gen2act/

Skanda Koppula, Ignacio Rocco, Yi Yang et al.

We introduce a new benchmark, TAPVid-3D, for evaluating the task of long-range Tracking Any Point in 3D (TAP-3D). While point tracking in two dimensions (TAP) has many benchmarks measuring performance on real-world videos, such as TAPVid-DAVIS, three-dimensional point tracking has none. To this end, leveraging existing footage, we build a new benchmark for 3D point tracking featuring 4,000+ real-world videos, composed of three different data sources spanning a variety of object types, motion patterns, and indoor and outdoor environments. To measure performance on the TAP-3D task, we formulate a collection of metrics that extend the Jaccard-based metric used in TAP to handle the complexities of ambiguous depth scales across models, occlusions, and multi-track spatio-temporal smoothness. We manually verify a large sample of trajectories to ensure correct video annotations, and assess the current state of the TAP-3D task by constructing competitive baselines using existing tracking models. We anticipate this benchmark will serve as a guidepost to improve our ability to understand precise 3D motion and surface deformation from monocular video. Code for dataset download, generation, and model evaluation is available at https://tapvid3d.github.io

Sebastian Jung, Artem Zholus, Martin Sundermeyer et al.

Tracking-Any-Point (TAP) models aim to track any point through a video which is a crucial task in AR/XR and robotics applications. The recently introduced TAPNext approach proposes an end-to-end, recurrent transformer architecture to track points frame-by-frame in a purely online fashion -- demonstrating competitive performance at minimal latency. However, we show that TAPNext struggles with longer video sequences and also frequently fails to re-detect query points that reappear after being occluded or leaving the frame. In this work, we present TAPNext++, a model that tracks points in sequences that are orders of magnitude longer while preserving the low memory and compute footprint of the architecture. We train the recurrent video transformer using several data-driven solutions, including training on long 1024-frame sequences enabled by sequence parallelism techniques. We highlight that re-detection performance is a blind spot in the current literature and introduce a new metric, Re-Detection Average Jaccard ($AJ_{RD}$), to explicitly evaluate tracking on re-appearing points. To improve re-detection of points, we introduce tailored geometric augmentations, such as periodic roll that simulates point re-entries, and supervising occluded points. We demonstrate that recurrent transformers can be substantially improved for point tracking and set a new state-of-the-art on multiple benchmarks. Model and code can be found at https://tap-next-plus-plus.github.io.

João Carreira, Dilara Gokay, Michael King et al.

Scaling has not yet been convincingly demonstrated for pure self-supervised learning from video. However, prior work has focused evaluations on semantic-related tasks $\unicode{x2013}$ action classification, ImageNet classification, etc. In this paper we focus on evaluating self-supervised learning on non-semantic vision tasks that are more spatial (3D) and temporal (+1D = 4D), such as camera pose estimation, point and object tracking, and depth estimation. We show that by learning from very large video datasets, masked auto-encoding (MAE) with transformer video models actually scales, consistently improving performance on these 4D tasks, as model size increases from 20M all the way to the largest by far reported self-supervised video model $\unicode{x2013}$ 22B parameters. Rigorous apples-to-apples comparison with many recent image and video models demonstrates the benefits of scaling 4D representations. Pretrained models are available at https://github.com/google-deepmind/representations4d .

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Kelsey Allen, Carl Doersch, Guangyao Zhou et al.

A current limitation of video generative video models is that they generate plausible looking frames, but poor motion -- an issue that is not well captured by FVD and other popular methods for evaluating generated videos. Here we go beyond FVD by developing a metric which better measures plausible object interactions and motion. Our novel approach is based on auto-encoding point tracks and yields motion features that can be used to not only compare distributions of videos (as few as one generated and one ground truth, or as many as two datasets), but also for evaluating motion of single videos. We show that using point tracks instead of pixel reconstruction or action recognition features results in a metric which is markedly more sensitive to temporal distortions in synthetic data, and can predict human evaluations of temporal consistency and realism in generated videos obtained from open-source models better than a wide range of alternatives. We also show that by using a point track representation, we can spatiotemporally localize generative video inconsistencies, providing extra interpretability of generated video errors relative to prior work. An overview of the results and link to the code can be found on the project page: http://trajan-paper.github.io.

Viorica Pătrăucean, Lucas Smaira, Ankush Gupta et al.

We propose a novel multimodal video benchmark - the Perception Test - to evaluate the perception and reasoning skills of pre-trained multimodal models (e.g. Flamingo, SeViLA, or GPT-4). Compared to existing benchmarks that focus on computational tasks (e.g. classification, detection or tracking), the Perception Test focuses on skills (Memory, Abstraction, Physics, Semantics) and types of reasoning (descriptive, explanatory, predictive, counterfactual) across video, audio, and text modalities, to provide a comprehensive and efficient evaluation tool. The benchmark probes pre-trained models for their transfer capabilities, in a zero-shot / few-shot or limited finetuning regime. For these purposes, the Perception Test introduces 11.6k real-world videos, 23s average length, designed to show perceptually interesting situations, filmed by around 100 participants worldwide. The videos are densely annotated with six types of labels (multiple-choice and grounded video question-answers, object and point tracks, temporal action and sound segments), enabling both language and non-language evaluations. The fine-tuning and validation splits of the benchmark are publicly available (CC-BY license), in addition to a challenge server with a held-out test split. Human baseline results compared to state-of-the-art video QA models show a substantial gap in performance (91.4% vs 46.2%), suggesting that there is significant room for improvement in multimodal video understanding. Dataset, baseline code, and challenge server are available at https://github.com/deepmind/perception_test

Carl Doersch, Pauline Luc, Yi Yang et al.

To endow models with greater understanding of physics and motion, it is useful to enable them to perceive how solid surfaces move and deform in real scenes. This can be formalized as Tracking-Any-Point (TAP), which requires the algorithm to track any point on solid surfaces in a video, potentially densely in space and time. Large-scale groundtruth training data for TAP is only available in simulation, which currently has a limited variety of objects and motion. In this work, we demonstrate how large-scale, unlabeled, uncurated real-world data can improve a TAP model with minimal architectural changes, using a selfsupervised student-teacher setup. We demonstrate state-of-the-art performance on the TAP-Vid benchmark surpassing previous results by a wide margin: for example, TAP-Vid-DAVIS performance improves from 61.3% to 67.4%, and TAP-Vid-Kinetics from 57.2% to 62.5%. For visualizations, see our project webpage at https://bootstap.github.io/

Daniel Geng, Charles Herrmann, Junhwa Hur et al.

Motion control is crucial for generating expressive and compelling video content; however, most existing video generation models rely mainly on text prompts for control, which struggle to capture the nuances of dynamic actions and temporal compositions. To this end, we train a video generation model conditioned on spatio-temporally sparse or dense motion trajectories. In contrast to prior motion conditioning work, this flexible representation can encode any number of trajectories, object-specific or global scene motion, and temporally sparse motion; due to its flexibility we refer to this conditioning as motion prompts. While users may directly specify sparse trajectories, we also show how to translate high-level user requests into detailed, semi-dense motion prompts, a process we term motion prompt expansion. We demonstrate the versatility of our approach through various applications, including camera and object motion control, "interacting" with an image, motion transfer, and image editing. Our results showcase emergent behaviors, such as realistic physics, suggesting the potential of motion prompts for probing video models and interacting with future generative world models. Finally, we evaluate quantitatively, conduct a human study, and demonstrate strong performance. Video results are available on our webpage: https://motion-prompting.github.io/

Artem Zholus, Carl Doersch, Yi Yang et al.

Tracking Any Point (TAP) in a video is a challenging computer vision problem with many demonstrated applications in robotics, video editing, and 3D reconstruction. Existing methods for TAP rely heavily on complex tracking-specific inductive biases and heuristics, limiting their generality and potential for scaling. To address these challenges, we present TAPNext, a new approach that casts TAP as sequential masked token decoding. Our model is causal, tracks in a purely online fashion, and removes tracking-specific inductive biases. This enables TAPNext to run with minimal latency, and removes the temporal windowing required by many existing state of art trackers. Despite its simplicity, TAPNext achieves a new state-of-the-art tracking performance among both online and offline trackers. Finally, we present evidence that many widely used tracking heuristics emerge naturally in TAPNext through end-to-end training. The TAPNext model and code can be found at https://tap-next.github.io/.

Sjoerd van Steenkiste, Daniel Zoran, Yi Yang et al.

Current vision models typically maintain a fixed correspondence between their representation structure and image space. Each layer comprises a set of tokens arranged "on-the-grid," which biases patches or tokens to encode information at a specific spatio(-temporal) location. In this work we present Moving Off-the-Grid (MooG), a self-supervised video representation model that offers an alternative approach, allowing tokens to move "off-the-grid" to better enable them to represent scene elements consistently, even as they move across the image plane through time. By using a combination of cross-attention and positional embeddings we disentangle the representation structure and image structure. We find that a simple self-supervised objective--next frame prediction--trained on video data, results in a set of latent tokens which bind to specific scene structures and track them as they move. We demonstrate the usefulness of MooG's learned representation both qualitatively and quantitatively by training readouts on top of the learned representation on a variety of downstream tasks. We show that MooG can provide a strong foundation for different vision tasks when compared to "on-the-grid" baselines.

Neerja Thakkar, Shiry Ginosar, Jacob Walker et al.

Visual intelligence requires anticipating the future behavior of agents, yet vision systems lack a general representation for motion and behavior. We propose dense point trajectories as visual tokens for behavior, a structured mid-level representation that disentangles motion from appearance and generalizes across diverse non-rigid agents, such as animals in-the-wild. Building on this abstraction, we design a diffusion transformer that models unordered sets of trajectories and explicitly reasons about occlusion, enabling coherent forecasts of complex motion patterns. To evaluate at scale, we curate 300 hours of unconstrained animal video with robust shot detection and camera-motion compensation. Experiments show that forecasting trajectory tokens achieves category-agnostic, data-efficient prediction, outperforms state-of-the-art baselines, and generalizes to rare species and morphologies, providing a foundation for predictive visual intelligence in the wild.

Jacob C Walker, Pedro Vélez, Luisa Polania Cabrera et al.

Forecasting what will happen next is a critical skill for general-purpose systems that plan or act in the world at different levels of abstraction. In this paper, we identify a strong correlation between a vision model's perceptual ability and its generalist forecasting performance over short time horizons. This trend holds across a diverse set of pretrained models-including those trained generatively-and across multiple levels of abstraction, from raw pixels to depth, point tracks, and object motion. The result is made possible by a novel generalist forecasting framework that operates on any frozen vision backbone: we train latent diffusion models to forecast future features in the frozen representation space, which are then decoded via lightweight, task-specific readouts. To enable consistent evaluation across tasks, we introduce distributional metrics that compare distributional properties directly in the space of downstream tasks and apply this framework to nine models and four tasks. Our results highlight the value of bridging representation learning and generative modeling for temporally grounded video understanding.

Wang Yifan, Carl Doersch, Relja Arandjelović et al.

Much of the recent progress in 3D vision has been driven by the development of specialized architectures that incorporate geometrical inductive biases. In this paper we tackle 3D reconstruction using a domain agnostic architecture and study how instead to inject the same type of inductive biases directly as extra inputs to the model. This approach makes it possible to apply existing general models, such as Perceivers, on this rich domain, without the need for architectural changes, while simultaneously maintaining data efficiency of bespoke models. In particular we study how to encode cameras, projective ray incidence and epipolar geometry as model inputs, and demonstrate competitive multi-view depth estimation performance on multiple benchmarks.

Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac et al.

A central goal of machine learning is the development of systems that can solve many problems in as many data domains as possible. Current architectures, however, cannot be applied beyond a small set of stereotyped settings, as they bake in domain & task assumptions or scale poorly to large inputs or outputs. In this work, we propose Perceiver IO, a general-purpose architecture that handles data from arbitrary settings while scaling linearly with the size of inputs and outputs. Our model augments the Perceiver with a flexible querying mechanism that enables outputs of various sizes and semantics, doing away with the need for task-specific architecture engineering. The same architecture achieves strong results on tasks spanning natural language and visual understanding, multi-task and multi-modal reasoning, and StarCraft II. As highlights, Perceiver IO outperforms a Transformer-based BERT baseline on the GLUE language benchmark despite removing input tokenization and achieves state-of-the-art performance on Sintel optical flow estimation with no explicit mechanisms for multiscale correspondence.

Carl Doersch, Ankush Gupta, Andrew Zisserman

Given new tasks with very little data$-$such as new classes in a classification problem or a domain shift in the input$-$performance of modern vision systems degrades remarkably quickly. In this work, we illustrate how the neural network representations which underpin modern vision systems are subject to supervision collapse, whereby they lose any information that is not necessary for performing the training task, including information that may be necessary for transfer to new tasks or domains. We then propose two methods to mitigate this problem. First, we employ self-supervised learning to encourage general-purpose features that transfer better. Second, we propose a novel Transformer based neural network architecture called CrossTransformers, which can take a small number of labeled images and an unlabeled query, find coarse spatial correspondence between the query and the labeled images, and then infer class membership by computing distances between spatially-corresponding features. The result is a classifier that is more robust to task and domain shift, which we demonstrate via state-of-the-art performance on Meta-Dataset, a recent dataset for evaluating transfer from ImageNet to many other vision datasets.

Jean-Bastien Grill, Florian Strub, Florent Altché et al.

We introduce Bootstrap Your Own Latent (BYOL), a new approach to self-supervised image representation learning. BYOL relies on two neural networks, referred to as online and target networks, that interact and learn from each other. From an augmented view of an image, we train the online network to predict the target network representation of the same image under a different augmented view. At the same time, we update the target network with a slow-moving average of the online network. While state-of-the art methods rely on negative pairs, BYOL achieves a new state of the art without them. BYOL reaches $74.3\%$ top-1 classification accuracy on ImageNet using a linear evaluation with a ResNet-50 architecture and $79.6\%$ with a larger ResNet. We show that BYOL performs on par or better than the current state of the art on both transfer and semi-supervised benchmarks. Our implementation and pretrained models are given on GitHub.

Carl Doersch, Andrew Zisserman

Synthetic visual data can provide practically infinite diversity and rich labels, while avoiding ethical issues with privacy and bias. However, for many tasks, current models trained on synthetic data generalize poorly to real data. The task of 3D human pose estimation is a particularly interesting example of this sim2real problem, because learning-based approaches perform reasonably well given real training data, yet labeled 3D poses are extremely difficult to obtain in the wild, limiting scalability. In this paper, we show that standard neural-network approaches, which perform poorly when trained on synthetic RGB images, can perform well when the data is pre-processed to extract cues about the person's motion, notably as optical flow and the motion of 2D keypoints. Therefore, our results suggest that motion can be a simple way to bridge a sim2real gap when video is available. We evaluate on the 3D Poses in the Wild dataset, the most challenging modern benchmark for 3D pose estimation, where we show full 3D mesh recovery that is on par with state-of-the-art methods trained on real 3D sequences, despite training only on synthetic humans from the SURREAL dataset.

Olivier J. Hénaff, Aravind Srinivas, Jeffrey De Fauw et al.

Human observers can learn to recognize new categories of images from a handful of examples, yet doing so with artificial ones remains an open challenge. We hypothesize that data-efficient recognition is enabled by representations which make the variability in natural signals more predictable. We therefore revisit and improve Contrastive Predictive Coding, an unsupervised objective for learning such representations. This new implementation produces features which support state-of-the-art linear classification accuracy on the ImageNet dataset. When used as input for non-linear classification with deep neural networks, this representation allows us to use 2-5x less labels than classifiers trained directly on image pixels. Finally, this unsupervised representation substantially improves transfer learning to object detection on the PASCAL VOC dataset, surpassing fully supervised pre-trained ImageNet classifiers.

Anurag Arnab, Carl Doersch, Andrew Zisserman

We present a bundle-adjustment-based algorithm for recovering accurate 3D human pose and meshes from monocular videos. Unlike previous algorithms which operate on single frames, we show that reconstructing a person over an entire sequence gives extra constraints that can resolve ambiguities. This is because videos often give multiple views of a person, yet the overall body shape does not change and 3D positions vary slowly. Our method improves not only on standard mocap-based datasets like Human 3.6M -- where we show quantitative improvements -- but also on challenging in-the-wild datasets such as Kinetics. Building upon our algorithm, we present a new dataset of more than 3 million frames of YouTube videos from Kinetics with automatically generated 3D poses and meshes. We show that retraining a single-frame 3D pose estimator on this data improves accuracy on both real-world and mocap data by evaluating on the 3DPW and HumanEVA datasets.

Victor Bapst, Alvaro Sanchez-Gonzalez, Carl Doersch et al.

Physical construction---the ability to compose objects, subject to physical dynamics, to serve some function---is fundamental to human intelligence. We introduce a suite of challenging physical construction tasks inspired by how children play with blocks, such as matching a target configuration, stacking blocks to connect objects together, and creating shelter-like structures over target objects. We examine how a range of deep reinforcement learning agents fare on these challenges, and introduce several new approaches which provide superior performance. Our results show that agents which use structured representations (e.g., objects and scene graphs) and structured policies (e.g., object-centric actions) outperform those which use less structured representations, and generalize better beyond their training when asked to reason about larger scenes. Model-based agents which use Monte-Carlo Tree Search also outperform strictly model-free agents in our most challenging construction problems. We conclude that approaches which combine structured representations and reasoning with powerful learning are a key path toward agents that possess rich intuitive physics, scene understanding, and planning.

Rohit Girdhar, João Carreira, Carl Doersch et al.

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 using only raw RGB frames as input.

Mateusz Malinowski, Carl Doersch

Visual QA is a pivotal challenge for higher-level reasoning, requiring understanding language, vision, and relationships between many objects in a scene. Although datasets like CLEVR are designed to be unsolvable without such complex relational reasoning, some surprisingly simple feed-forward, "holistic" models have recently shown strong performance on this dataset. These models lack any kind of explicit iterative, symbolic reasoning procedure, which are hypothesized to be necessary for counting objects, narrowing down the set of relevant objects based on several attributes, etc. The reason for this strong performance is poorly understood. Hence, our work analyzes such models, and finds that minor architectural elements are crucial to performance. In particular, we find that \textit{early fusion} of language and vision provides large performance improvements. This contrasts with the late fusion approaches popular at the dawn of Visual QA. We propose a simple module we call Multimodal Core, which we hypothesize performs the fundamental operations for multimodal tasks. We believe that understanding why these elements are so important to complex question answering will aid the design of better-performing algorithms for Visual QA while minimizing hand-engineering effort.

Mateusz Malinowski, Carl Doersch, Adam Santoro et al.

Attention mechanisms in biological perception are thought to select subsets of perceptual information for more sophisticated processing which would be prohibitive to perform on all sensory inputs. In computer vision, however, there has been relatively little exploration of hard attention, where some information is selectively ignored, in spite of the success of soft attention, where information is re-weighted and aggregated, but never filtered out. Here, we introduce a new approach for hard attention and find it achieves very competitive performance on a recently-released visual question answering datasets, equalling and in some cases surpassing similar soft attention architectures while entirely ignoring some features. Even though the hard attention mechanism is thought to be non-differentiable, we found that the feature magnitudes correlate with semantic relevance, and provide a useful signal for our mechanism's attentional selection criterion. Because hard attention selects important features of the input information, it can also be more efficient than analogous soft attention mechanisms. This is especially important for recent approaches that use non-local pairwise operations, whereby computational and memory costs are quadratic in the size of the set of features.

Rohit Girdhar, João Carreira, Carl Doersch et al.

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.

Simon Schmitt, Jonathan J. Hudson, Augustin Zidek et al.

We present a method for using previously-trained 'teacher' agents to kickstart the training of a new 'student' agent. To this end, we leverage ideas from policy distillation and population based training. Our method places no constraints on the architecture of the teacher or student agents, and it regulates itself to allow the students to surpass their teachers in performance. We show that, on a challenging and computationally-intensive multi-task benchmark (DMLab-30), kickstarted training improves the data efficiency of new agents, making it significantly easier to iterate on their design. We also show that the same kickstarting pipeline can allow a single student agent to leverage multiple 'expert' teachers which specialize on individual tasks. In this setting kickstarting yields surprisingly large gains, with the kickstarted agent matching the performance of an agent trained from scratch in almost 10x fewer steps, and surpassing its final performance by 42 percent. Kickstarting is conceptually simple and can easily be incorporated into reinforcement learning experiments.

Carl Doersch, Andrew Zisserman

We investigate methods for combining multiple self-supervised tasks--i.e., supervised tasks where data can be collected without manual labeling--in order to train a single visual representation. First, we provide an apples-to-apples comparison of four different self-supervised tasks using the very deep ResNet-101 architecture. We then combine tasks to jointly train a network. We also explore lasso regularization to encourage the network to factorize the information in its representation, and methods for "harmonizing" network inputs in order to learn a more unified representation. We evaluate all methods on ImageNet classification, PASCAL VOC detection, and NYU depth prediction. Our results show that deeper networks work better, and that combining tasks--even via a naive multi-head architecture--always improves performance. Our best joint network nearly matches the PASCAL performance of a model pre-trained on ImageNet classification, and matches the ImageNet network on NYU depth prediction.

Jacob Walker, Carl Doersch, Abhinav Gupta et al.

In a given scene, humans can often easily predict a set of immediate future events that might happen. However, generalized pixel-level anticipation in computer vision systems is difficult because machine learning struggles with the ambiguity inherent in predicting the future. In this paper, we focus on predicting the dense trajectory of pixels in a scene, specifically what will move in the scene, where it will travel, and how it will deform over the course of one second. We propose a conditional variational autoencoder as a solution to this problem. In this framework, direct inference from the image shapes the distribution of possible trajectories, while latent variables encode any necessary information that is not available in the image. We show that our method is able to successfully predict events in a wide variety of scenes and can produce multiple different predictions when the future is ambiguous. Our algorithm is trained on thousands of diverse, realistic videos and requires absolutely no human labeling. In addition to non-semantic action prediction, we find that our method learns a representation that is applicable to semantic vision tasks.

Carl Doersch

In just three years, Variational Autoencoders (VAEs) have emerged as one of the most popular approaches to unsupervised learning of complicated distributions. VAEs are appealing because they are built on top of standard function approximators (neural networks), and can be trained with stochastic gradient descent. VAEs have already shown promise in generating many kinds of complicated data, including handwritten digits, faces, house numbers, CIFAR images, physical models of scenes, segmentation, and predicting the future from static images. This tutorial introduces the intuitions behind VAEs, explains the mathematics behind them, and describes some empirical behavior. No prior knowledge of variational Bayesian methods is assumed.

Philipp Krähenbühl, Carl Doersch, Jeff Donahue et al.

Convolutional Neural Networks spread through computer vision like a wildfire, impacting almost all visual tasks imaginable. Despite this, few researchers dare to train their models from scratch. Most work builds on one of a handful of ImageNet pre-trained models, and fine-tunes or adapts these for specific tasks. This is in large part due to the difficulty of properly initializing these networks from scratch. A small miscalibration of the initial weights leads to vanishing or exploding gradients, as well as poor convergence properties. In this work we present a fast and simple data-dependent initialization procedure, that sets the weights of a network such that all units in the network train at roughly the same rate, avoiding vanishing or exploding gradients. Our initialization matches the current state-of-the-art unsupervised or self-supervised pre-training methods on standard computer vision tasks, such as image classification and object detection, while being roughly three orders of magnitude faster. When combined with pre-training methods, our initialization significantly outperforms prior work, narrowing the gap between supervised and unsupervised pre-training.

Carl Doersch, Abhinav Gupta, Alexei A. Efros

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

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.