Video temporal grounding aims to localize relevant temporal boundaries in a video given a textual prompt. Recent work has focused on enabling Video LLMs to perform video temporal grounding via next-token prediction of temporal timestamps. However, accurately localizing timestamps in videos remains challenging for Video LLMs when relying solely on temporal token prediction. Our proposed TimeRefine addresses this challenge in two ways. First, instead of directly predicting the start and end timestamps, we reformulate temporal grounding as a temporal refinement task: the model first makes rough predictions and then refines them by predicting offsets to the target segment. This refining process is repeated multiple times, through which the model progressively improves its own temporal localization accuracy. Second, to enhance the model’s temporal perception capabilities, we incorporate an auxiliary prediction head that applies a larger penalty as a predicted segment deviates further from the ground truth, encouraging more precise temporal localizations. Our plug-and-play method can be integrated into most LLM-based temporal grounding approaches. The experimental results demonstrate that TimeRefine achieves 3.6% and 5.0% mIoU improvements on the ActivityNet and Charades-STA datasets, respectively. Code and pretrained models are available at https://github.com/SJTUwxz/TimeRefine_code.

Video Question Answering (VQA) in long videos poses the key challenge of extracting relevant information and modeling long-range dependencies from many redundant frames. The self-attention mechanism provides a general solution for sequence modeling, but it has a prohibitive cost when applied to a massive number of spatiotemporal tokens in long videos. Most prior methods rely on compression strategies to lower the computational cost, such as reducing the input length via sparse frame sampling or compressing the output sequence passed to the large language model (LLM) via space-time pooling. However, these naive approaches over-represent redundant information and often miss salient events or fast-occurring space-time patterns. In this work, we introduce BIMBA, an efficient state-space model to handle long-form videos. Our model leverages the selective scan algorithm to learn to effectively select critical information from high-dimensional video and transform it into a reduced token sequence for efficient LLM processing. Extensive experiments demonstrate that BIMBA achieves state-of-the-art accuracy on multiple long-form VQA benchmarks, including PerceptionTest, NExT-QA, EgoSchema, VNBench, LongVideoBench, Video-MME, and MLVU. Code and models are available at https://sites.google.com/view/bimba-mllm.

We investigate complex video question answering via chain-of-evidence reasoning — identifying sequences of temporal spans from multiple relevant parts of the video, together with visual evidence within them. Existing models struggle with multi-step reasoning as they uniformly sample a fixed number of frames, which can miss critical evidence distributed nonuniformly throughout the video. Moreover, they lack the ability to temporally localize such evidence in the broader context of the full video, which is required for answering complex questions. We propose a framework to enhance existing VideoQA datasets with evidence reasoning chains, automatically constructed by searching for optimal intervals of interest in the video with supporting evidence, that maximizes the likelihood of answering a given question. We train our model (VITED) to generate these evidence chains directly, enabling it to both localize evidence windows as well as perform multi-step reasoning across them in long-form video content. We show the value of our evidence-distilled models on a suite of long video QA benchmarks where we outperform state-of-the-art approaches that lack evidence reasoning capabilities.

We introduce ED-VTG, a method for fine-grained video temporal grounding utilizing multi-modal large language models. Our approach harnesses the capabilities of multimodal LLMs to jointly process text and video, in order to effectively localize natural language queries in videos through a two-stage process. Rather than being directly grounded, language queries are initially transformed into enriched sentences that incorporate missing details and cues to aid in grounding. In the second stage, these enriched queries are grounded, using a lightweight decoder, which specializes at predicting accurate boundaries conditioned on contextualized representations of the enriched queries. To mitigate noise and reduce the impact of hallucinations, our model is trained with a multiple-instance-learning objective that dynamically selects the optimal version of the query for each training sample. We demonstrate state-of-the-art results across various benchmarks in temporal video grounding and paragraph grounding settings. Experiments reveal that our method significantly outperforms all previously proposed LLM-based temporal grounding approaches and is either superior or comparable to specialized models, while maintaining a clear advantage against them in zero-shot evaluation scenarios.

Vision-language models are integral to computer vision research, yet many high-performing models remain closed-source, obscuring their data, design and training recipe. The research community has responded by using distillation from black-box models to label training data, achieving strong benchmark results, at the cost of measurable scientific progress. However, without knowing the details of the teacher model and its data sources, scientific progress remains difficult to measure. In this paper, we study building a Perception Language Model (PLM) in a fully open and reproducible framework for transparent research in image and video understanding. We analyze standard training pipelines without distillation from proprietary models and explore large-scale synthetic data to identify critical data gaps, particularly in detailed video understanding. To bridge these gaps, we release 2.8M human-labeled instances of fine-grained video question-answer pairs and spatio-temporally grounded video captions. Additionally, we introduce PLM–VideoBench, a suite for evaluating challenging video understanding tasks focusing on the ability to reason about ''what'', ''where'', ''when'', and ''how'' of a video. We make our work fully reproducible by providing data, training recipes, code & models.

We present Ego-Exo4D, a diverse, large-scale multimodal multiview video dataset and benchmark challenge. Ego-Exo4D centers around simultaneously-captured egocentric and exocentric video of skilled human activities (e.g., sports, music, dance, bike repair). 740 participants from 13 cities worldwide performed these activities in 123 different natural scene contexts, yielding long-form captures from 1 to 42 minutes each and 1,286 hours of video combined. The multimodal nature of the dataset is unprecedented: the video is accompanied by multichannel audio, eye gaze, 3D point clouds, camera poses, IMU, and multiple paired language descriptions—including a novel “expert commentary” done by coaches and teachers and tailored to the skilled-activity domain. To push the frontier of first-person video understanding of skilled human activity, we also present a suite of benchmark tasks and their annotations, including fine-grained activity understanding, proficiency estimation, cross-view translation, and 3D hand/body pose. All resources are open sourced to fuel new research in the community. https://ego-exo4d-data.org/

We introduce Ego4D, a massive-scale egocentric video dataset and benchmark suite. It offers 3,670 hours of daily-life activity video spanning hundreds of scenarios (household, outdoor, workplace, leisure, etc.) captured by 931 unique camera wearers from 74 worldwide locations and 9 different countries. The approach to collection is designed to uphold rigorous privacy and ethics standards, with consenting participants and robust de-identification procedures where relevant. Ego4D dramatically expands the volume of diverse egocentric video footage publicly available to the research community. Portions of the video are accompanied by audio, 3D meshes of the environment, eye gaze, stereo, and/or synchronized videos from multiple egocentric cameras at the same event. Furthermore, we present a host of new benchmark challenges centered around understanding the first-person visual experience in the past (querying an episodic memory), present (analyzing hand-object manipulation, audio-visual conversation, and social interactions), and future (forecasting activities). By publicly sharing this massive annotated dataset and benchmark suite, we aim to push the frontier of first-person perception.

We present Ego-Exo4D, a diverse, large-scale multi-modal multiview video dataset and benchmark challenge. Ego-Exo4D centers around simultaneously-captured ego-centric and exocentric video of skilled human activities (e.g., sports, music, dance, bike repair). 740 participants from 13 cities worldwide performed these activities in 123 different natural scene contexts, yielding long-form captures from 1 to 42 minutes each and 1,286 hours of video combined. The multimodal nature of the dataset is un-precedented: the video is accompanied by multichannel audio, eye gaze, 3D point clouds, camera poses, IMU, and multiple paired language descriptions-including a novel “expert commentary” done by coaches and teachers and tailored to the skilled-activity domain. To push the frontier of first-person video understanding of skilled human activity, we also present a suite of benchmark tasks and their annotations, including fine-grained activity understanding, proficiency estimation, cross-view translation, and 3D hand/body pose. All resources are open sourced to fuel new research in the community.

Most video captioning models are designed to process short video clips of few seconds and output text describing low-level visual concepts (e.g., objects, scenes, atomic actions). However, most real-world videos last for minutes or hours and have a complex hierarchical structure spanning different temporal granularities. We propose Video ReCap, a recursive video captioning model that can process video inputs of dramatically different lengths (from 1 second to 2 hours) and output video captions at multiple hierarchy levels. The recursive video-language architecture exploits the synergy between different video hierarchies and can process hour-long videos efficiently. We utilize a curriculum learning training scheme to learn the hierarchical structure of videos, starting from clip-level captions describing atomic actions, then focusing on segment-level descriptions, and concluding with generating summaries for hour-long videos. Furthermore, we introduce Ego4D-HCap dataset by augmenting Ego4D with 8,267 manually collected long-range video summaries. Our recursive model can flexibly generate captions at different hierarchy levels while also being useful for other complex video understanding tasks, such as VideoQA on EgoSchema. Data, code, and models are publicly available at https://sites.google.com/view/vidrecap.

Comparing a user video to a reference how-to video is a key requirement for AR/VR technology delivering personalized assistance tailored to the user's progress. However, current approaches for language-based assistance can only answer questions about a single video. We propose an approach that first automatically generates large amounts of visual instruction tuning data involving pairs of videos from HowTo100M by leveraging existing step annotations and accompanying narrations, and then trains a video-conditioned language model to jointly reason across multiple raw videos. Our model achieves state-of-the-art performance at identifying differences between video pairs and ranking videos based on the severity of these differences, and shows promising ability to perform general reasoning over multiple videos.

Egocentric videos provide a first-person perspective of the wearer's activities, involving simultaneous interactions with multiple objects. In this work, we propose the task of weakly-supervised Narration-based Video Object Segmentation (NVOS). Given an egocentric video clip and a narration of the wearer's activities, our aim is to segment object instances mentioned in the narration, with-out using any spatial annotations during training. Existing weakly-supervised video object grounding methods typ-ically yield bounding boxes for referred objects. In contrast, we propose ROSA, a weakly-supervised pixel-level grounding framework learning alignments between referred objects and segmentation mask proposals. Our model harnesses vision-language models pre-trained on image-text pairs to embed region masks and object phrases. During training, we combine (a) a video-narration contrastive loss that implicitly supervises the alignment between regions and phrases, and (b) a region-phrase contrastive loss based on inferred latent alignments. To address the lack of annotated NVOS datasets in egocentric videos, we create a new evaluation benchmark, VISOR-NVOS, leveraging existing annotations of segmentation masks from VISOR alongside 14.6k newly-collected, object-based video clip narrations. Our approach achieves state-of-the-art zero-shot pixel-level grounding performance compared to strong baselines under similar supervision. Additionally, we demonstrate generalization capabilities for zero-shot video object grounding on YouCook2, a third-person instructional video dataset.

Yuanhao Xiong, Yixin Nie, Haotian Liu, Boxin Wang, Jun Chen, Rong Jin, Cho-Jui Hsieh, Lorenzo Torresani, Jie Lei. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Video-language embeddings are a promising avenue for injecting semantics into visual representations, but existing methods capture only short-term associations between seconds-long video clips and their accompanying text. We propose HierVL, a novel hierarchical video-language em-bedding that simultaneously accounts for both long-term and short-term associations. As training data, we take videos accompanied by timestamped text descriptions of human actions, together with a high-level text summary of the activity throughout the long video (as are available in Ego4D). We introduce a hierarchical contrastive training objective that encourages text-visual alignment at both the clip level and video level. While the clip-level constraints use the step-by-step descriptions to capture what is happening in that instant, the video-level constraints use the summary text to capture why it is happening, i.e., the broader context for the activity and the intent of the actor. Our hierarchical scheme yields a clip representation that outperforms its single-level counterpart as well as a long-term video representation that achieves SotA results on tasks requiring long-term video modeling. HierVL success-fully transfers to multiple challenging downstream tasks (in EPIC-KITCHENS-100, Charades-Ego, HowTo100M) in both zero-shot and fine-tuned settings.

Different video understanding tasks are typically treated in isolation, and even with distinct types of curated data (e.g., classifying sports in one dataset, tracking animals in another). However, in wearable cameras, the immersive egocentric perspective of a person engaging with the world around them presents an interconnected web of video understanding tasks—hand-object manipulations, navigation in the space, or human-human interactions—that unfold continuously, driven by the person's goals. We argue that this calls for a much more unified approach. We propose EgoTask Translation (EgoT2), which takes a collection of models optimized on separate tasks and learns to translate their outputs for improved performance on any or all of them at once. Unlike traditional transfer or multi-task learning, EgoT2's “flipped design” entails separate task-specific backbones and a task translator shared across all tasks, which captures synergies between even heterogeneous tasks and mitigates task competition. Demonstrating our model on a wide array of video tasks from Ego4D, we show its advantages over existing transfer paradigms and achieve top-ranked results on four of the Ego4D 2022 benchmark challenges.11Project webpage: https://vision.cs.utexas.edu/projects/egot2/.

Egocentric videos are often available in the form of uninterrupted, uncurated long videos capturing the camera wearers' daily life activities. Understanding these videos requires models to be able to reason about activities, objects, and their interactions. However, current video benchmarks study these problems independently and under short, curated clips. In contrast, real-world applications, e.g. AR assistants, require bundling these problems for both model development and evaluation. In this paper, we propose to study these problems in a joint framework for long video understanding. Our contributions are three-fold. First, we propose an integrated framework, namely Relational Space-Time Query (ReST), for evaluating video understanding models via templated spatiotemporal queries. Second, we introduce two new benchmarks, ReST-ADL and ReST-Ego4D11The latest version of our benchmark and models will be available here., which augment the existing egocentric video datasets with abundant query annotations generated by the ReST framework. Finally, we present a set of baselines and in-depth analysis on the two benchmarks and provide insights about the query tasks. We view our integrated framework and benchmarks as a step towards comprehensive, multi-step reasoning in long videos, and believe it will facilitate the development of next generations of video understanding models.

In this paper we present an approach for localizing steps of procedural activities in narrated how-to videos. To deal with the scarcity of labeled data at scale, we source the step descriptions from a language knowledge base (wikiHow) containing instructional articles for a large variety of procedural tasks. Without any form of manual supervision, our model learns to temporally ground the steps of procedural articles in how-to videos by matching three modalities: frames, narrations, and step descriptions. Specifically, our method aligns steps to video by fusing information from two distinct pathways: i) direct alignment of step descriptions to frames, ii) indirect alignment obtained by composing steps-to-narrations with narrations-to-video correspondences. Notably, our approach performs global temporal grounding of all steps in an article at once by exploiting order information, and is trained with step pseudo-labels which are iteratively refined and aggressively filtered. In order to validate our model we introduce a new benchmark – HT-Step – obtained by manually annotating a 124-hour subset of HowTo100M1with steps sourced from wikiHow articles. Experiments on this benchmark as well as zero-shot evaluations on CrossTask demonstrate that our multi-modality alignment yields dramatic gains over several baselines and prior works. Finally, we show that our inner module for matching narration-to-video outperforms by a large margin the state of the art on the HTM-Align narration-video alignment benchmark.

We present Ego-Only, the first approach that enables state-of-the-art action detection on egocentric (first-person) videos without any form of exocentric (third-person) transferring. Despite the content and appearance gap separating the two domains, large-scale exocentric transferring has been the default choice for egocentric action detection. This is because prior works found that egocentric models are difficult to train from scratch and that transferring from exocentric representations leads to improved accuracy. However, in this paper, we revisit this common belief. Motivated by the large gap separating the two domains, we propose a strategy that enables effective training of egocentric models without exocentric transferring. Our Ego-Only approach is simple. It trains the video representation with a masked autoencoder finetuned for temporal segmentation. The learned features are then fed to an off-the-shelf temporal action localization method to detect actions. We find that this renders exocentric transferring unnecessary by showing remarkably strong results achieved by this simple Ego-Only approach on three established egocentric video datasets: Ego4D, EPIC-Kitchens-100, and Charades-Ego. On both action detection and action recognition, Ego-Only outperforms previous best exocentric transferring methods that use orders of magnitude more labels. Ego-Only sets new state-of-the-art results on these datasets and benchmarks without exocentric data.

We introduce HT-Step, a large-scale dataset containing temporal annotations of instructional article steps in cooking videos. It includes 122k segment-level annotations over 20k narrated videos (approximately 2.3k hours) of the HowTo100M dataset.Each annotation provides a temporal interval, and a categorical step label from a taxonomy of 4,958 unique steps automatically mined from wikiHow articles which include rich descriptions of each step.Our dataset significantly surpasses existing labeled step datasets in terms of scale, number of tasks, and richness of natural language step descriptions. Based on these annotations, we introduce a strongly supervised benchmark for aligning instructional articles with how-to videos and present a comprehensive evaluation of baseline methods for this task.By publicly releasing these annotations and defining rigorous evaluation protocols and metrics,we hope to significantly accelerate research in the field of procedural activity understanding.

Human activities are goal-oriented and hierarchical, comprising primary goals at the top level, sequences of steps and substeps in the middle, and atomic actions at the lowest level. Recognizing human activities thus requires relating atomic actions and steps to their functional objectives (what the actions contribute to) and modeling their sequential and hierarchical dependencies towards achieving the goals. Current activity recognition research has primarily focused on only the lowest levels of this hierarchy, i.e., atomic or low-level actions, often in trimmed videos with annotations spanning only a few seconds. In this work, we introduce Ego4D Goal-Step, a new set of annotations on the recently released Ego4D with a novel hierarchical taxonomy of goal-oriented activity labels. It provides dense annotations for 48K procedural step segments (430 hours) and high-level goal annotations for 2,807 hours of Ego4D videos. Compared to existing procedural video datasets, it is substantially larger in size, contains hierarchical action labels (goals - steps - substeps), and provides goal-oriented auxiliary information including natural language summary description, step completion status, and step-to-goal relevance information. We take a data-driven approach to build our taxonomy, resulting in dense step annotations that do not suffer from poor label-data alignment issues resulting from a taxonomy defined a priori. Through comprehensive evaluations and analyses, we demonstrate how Ego4D Goal-Step supports exploring various questions in procedural activity understanding, including goal inference, step prediction, hierarchical relation learning, and long-term temporal modeling.

Few-shot classification requires adapting knowledge learned from a large annotated base dataset to recognize novel unseen classes, each represented by few labeled examples. In such a scenario, pretraining a network with high capacity on the large dataset and then finetuning it on the few examples causes severe overfitting. At the same time, training a simple linear classifier on top of ``frozen'' features learned from the large labeled dataset fails to adapt the model to the properties of the novel classes, effectively inducing underfitting. In this paper we propose an alternative approach to both of these two popular strategies. First, our method pseudo-labels the entire large dataset using the linear classifier trained on the novel classes. This effectively ``hallucinates'' the novel classes in the large dataset, despite the novel categories not being present in the base database (novel and base classes are disjoint). Then, it finetunes the entire model with a distillation loss on the pseudo-labeled base examples, in addition to the standard cross-entropy loss on the novel dataset. This step effectively trains the network to recognize contextual and appearance cues that are useful for the novel-category recognition but using the entire large-scale base dataset and thus overcoming the inherent data-scarcity problem of few-shot learning. Despite the simplicity of the approach, we show that that our method outperforms the state-of-the-art on four well-established few-shot classification benchmarks. The code is available at https://github.com/yiren-jian/LabelHalluc.

In this paper we consider the problem of classifying fine-grained, multi-step activities (e.g., cooking different recipes, making disparate home improvements, creating various forms of arts and crafts) from long videos spanning up to several minutes. Accurately categorizing these activities requires not only recognizing the individual steps that compose the task but also capturing their temporal dependencies. This problem is dramatically different from traditional action classification, where models are typically optimized on videos that span only a few seconds and that are manually trimmed to contain simple atomic actions. While step annotations could enable the training of models to recognize the individual steps of procedural activities, existing large-scale datasets in this area do not include such segment labels due to the prohibitive cost of manually annotating temporal boundaries in long videos. To address this issue, we propose to automatically identify steps in instructional videos by leveraging the distant supervision of a textual knowledge base (wikiHow) that includes detailed descriptions of the steps needed for the execution of a wide variety of complex activities. Our method uses a language model to match noisy, automatically-transcribed speech from the video to step descriptions in the knowledge base. We demonstrate that video models trained to recognize these automatically-labeled steps (without manual supervision) yield a representation that achieves superior generalization performance on four downstream tasks: recognition of procedural activities, step classification, step forecasting and egocentric video classification.

We introduce Ego4D, a massive-scale egocentric video dataset and benchmark suite. It offers 3,670 hours of dailylife activity video spanning hundreds of scenarios (household, outdoor, workplace, leisure, etc.) captured by 931 unique camera wearers from 74 worldwide locations and 9 different countries. The approach to collection is designed to uphold rigorous privacy and ethics standards, with consenting participants and robust de-identification procedures where relevant. Ego4D dramatically expands the volume of diverse egocentric video footage publicly available to the research community. Portions of the video are accompanied by audio, 3D meshes of the environment, eye gaze, stereo, and/or synchronized videos from multiple egocentric cameras at the same event. Furthermore, we present a host of new benchmark challenges centered around understanding the first-person visual experience in the past (querying an episodic memory), present (analyzing hand-object manipulation, audio-visual conversation, and social interactions), and future (forecasting activities). By publicly sharing this massive annotated dataset and benchmark suite, we aim to push the frontier of first-person perception. Project page: https://ego4d-data.org/

Video transformers have recently emerged as a competitive alternative to 3D CNNs for video understanding. However, due to their large number of parameters and reduced inductive biases, these models require supervised pretraining on large-scale image datasets to achieve top performance. In this paper, we empirically demonstrate that self-supervised pretraining of video transformers on video-only datasets can lead to action recognition results that are on par or better than those obtained with supervised pretraining on large-scale image datasets, even massive ones such as ImageNet-21K. Since transformer-based models are effective at capturing dependencies over extended temporal spans, we propose a simple learning procedure that forces the model to match a long-term view to a short-term view of the same video. Our approach, named Long-Short Temporal Contrastive Learning (LSTCL), enables video transformers to learn an effective clip-level representation by predicting temporal context captured from a longer temporal extent. To demonstrate the generality of our findings, we implement and validate our approach under three different self-supervised contrastive learning frameworks (MoCo v3, BYOL, SimSiam) using two distinct video-transformer architectures, including an improved variant of the Swin Transformer augmented with space-time attention. We conduct a thorough ablation study and show that LSTCL achieves competitive performance on multiple video benchmarks and represents a convincing alternative to supervised image-based pretraining.

Video transformers have recently emerged as an effective alternative to convolutional networks for action classification. However, most prior video transformers adopt either global space-time attention or hand-defined strategies to compare patches within and across frames. These fixed attention schemes not only have high computational cost but, by comparing patches at predetermined locations, they neglect the motion dynamics in the video. In this paper, we introduce the Deformable Video Transformer (DVT), which dynamically predicts a small subset of video patches to attend for each query location based on motion information, thus allowing the model to decide where to look in the video based on correspondences across frames. Crucially, these motion-based correspondences are obtained at zero-cost from information stored in the compressed format of the video. Our deformable attention mechanism is optimized directly with respect to classification performance, thus eliminating the need for suboptimal hand-design of attention strategies. Experiments on four large-scale video benchmarks (Kinetics-400, Something-Something-V2, EPIC-KITCHENS and Diving-48) demonstrate that, compared to existing video transformers, our model achieves higher accuracy at the same or lower computational cost, and it attains state-of-the-art results on these four datasets.

Few-shot learning aims to recognize novel classes from a few examples. Although significant progress has been made in the image domain, few-shot video classification is relatively unexplored. We argue that previous methods underestimate the importance of video feature learning and propose to learn spatiotemporal features using a 3D CNN. Proposing a two-stage approach that learns video features on base classes followed by fine-tuning the classifiers on novel classes, we show that this simple baseline approach outperforms prior few-shot video classification methods by over 20 points on existing benchmarks. To circumvent the need of labeled examples, we present two novel approaches that yield further improvement. First, we leverage tag-labeled videos from a large dataset using tag retrieval followed by selecting the best clips with visual similarities. Second, we learn generative adversarial networks that generate video features of novel classes from their semantic embeddings. Moreover, we find existing benchmarks are limited because they only focus on 5 novel classes in each testing episode and introduce more realistic benchmarks by involving more novel classes, i.e., few-shot learning, as well as a mixture of novel and base classes, i.e., generalized few-shot learning. The experimental results show that our retrieval and feature generation approach significantly outperform the baseline approach on the new benchmarks.

We present VX2TEXT, a framework for text generation from multimodal inputs consisting of video plus text, speech, or audio. In order to leverage transformer networks, which have been shown to be effective at modeling language, each modality is first converted into a set of language embeddings by a learnable tokenizer. This allows our approach to perform multimodal fusion in the language space, thus eliminating the need for ad-hoc cross-modal fusion modules. To address the non-differentiability of tokenization on continuous inputs (e.g., video or audio), we utilize a relaxation scheme that enables end-to-end training. Furthermore, unlike prior encoder-only models, our network includes an autoregressive decoder to generate open-ended text from the multimodal embeddings fused by the language encoder. This renders our approach fully generative and makes it directly applicable to different "video+x to text" problems without the need to design specialized network heads for each task. The proposed framework is not only conceptually simple but also remarkably effective: experiments demonstrate that our approach based on a single architecture outperforms the state-of-the-art on three videobased text-generation tasks—captioning, question answering and audio-visual scene-aware dialog.

The standard way of training video models entails sampling at each iteration a single clip from a video and optimizing the clip prediction with respect to the video-level label. We argue that a single clip may not have enough temporal coverage to exhibit the label to recognize, since video datasets are often weakly labeled with categorical information but without dense temporal annotations. Furthermore, optimizing the model over brief clips impedes its ability to learn long-term temporal dependencies. To overcome these limitations, we introduce a collaborative memory mechanism that encodes information across multiple sampled clips of a video at each training iteration. This enables the learning of long-range dependencies beyond a single clip. We explore different design choices for the collaborative memory to ease the optimization difficulties. Our proposed framework is end-to-end trainable and significantly improves the accuracy of video classification at a negligible computational overhead. Through extensive experiments, we demonstrate that our framework generalizes to different video architectures and tasks, outperforming the state of the art on both action recognition (e.g., Kinetics-400 & 700, Charades, Something-Something-V1) and action detection (e.g., AVA v2.1 & v2.2).

Although speaker verification has conventionally been an audio-only task, some practical applications provide both audio and visual streams of input. In these cases, the visual stream provides complementary information and can often be leveraged in conjunction with the acoustics of speech to improve verification performance. In this study, we explore audio-visual approaches to speaker verification, starting with standard fusion techniques to learn joint audio-visual (AV) embeddings, and then propose a novel approach to handle cross-modal verification at test time. Specifically, we investigate unimodal and concatenation based AV fusion and report the lowest AV equal error rate (EER) of 0.7% on the VoxCeleb1 dataset using our best system. As these methods lack the ability to do cross-modal verification, we introduce a multi-view model which uses a shared classifier to map audio and video into the same space. This new approach achieves 28% EER on VoxCeleb1 in the challenging testing condition of cross-modal verification.

In contrast to traditional weight optimization in a continuous space, we demonstrate the existence of effective random networks whose weights are never updated. By selecting a weight among a fixed set of random values for each individual connection, our method uncovers combinations of random weights that match the performance of traditionally-trained networks of the same capacity. We refer to our networks as "slot machines" where each reel (connection) contains a fixed set of symbols (random values). Our backpropagation algorithm "spins" the reels to seek "winning" combinations, i.e., selections of random weight values that minimize the given loss. Quite surprisingly, we find that allocating just a few random values to each connection (e.g., 8 values per connection) yields highly competitive combinations despite being dramatically more constrained compared to traditionally learned weights. Moreover, finetuning these combinations often improves performance over the trained baselines. A randomly initialized VGG-19 with 8 values per connection contains a combination that achieves 91% test accuracy on CIFAR-10. Our method also achieves an impressive performance of 98.2% on MNIST for neural networks containing only random weights.

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named “TimeSformer,” adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that “divided attention,” where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

Understanding temporal information and how the visual world changes over time, is a fundamental ability of intelligent systems. In video understanding, temporal information is at the core of many current challenges, including compression, efficient inference, motion estimation or summarization. However, in current video datasets it has been observed that action classes can often be recognized without any temporal information, from a single frame of video. As a result, both benchmarking and training in these datasets may give an unintentional advantage to models with strong image understanding capabilities, as opposed to those with strong temporal understanding. In other words, current datasets may not reward good temporal understanding, potentially hindering progress. In this paper we address this problem head on by identifying action classes where temporal information is actually necessary to recognize them and call these "temporal classes". Selecting temporal classes using a computational method would bias the process. Instead, we propose a methodology based on a simple and effective human annotation experiment. We remove just the temporal information, by shuffling frames in time, and measure if the action can still be recognized. Classes that cannot be recognized when frames are not in order, are included in the temporal set. We observe that this set is statistically different from other static classes, and that performance in it correlates with a network's ability to capture temporal information. Thus we use it as a benchmark on current popular networks, which reveals a series of interesting facts, like inflated convolutions bias networks towards classes where motion is not important. We also explore the effect of training on the temporal set, and observe that this leads to better generalization in unseen classes, demonstrating the need for more temporal data. We hope that the proposed dataset of temporal categories will help guide future research in temporal modeling for better video understanding.

A significant challenge in video understanding is posed by the high dimensionality of the input, which induces large computational cost and high memory footprints. Deep convolutional models operating on video apply pooling and striding to reduce feature dimensionality and to increase the receptive field. However, despite these strategies, modern approaches cannot effectively leverage spatiotemporal structure over long temporal extents. In this paper we introduce an approach that reduces a video of 10 seconds to a sparse graph of only 160 feature nodes such that efficient inference in this graph produces state-of-the-art accuracy on challenging action recognition datasets. The nodes of our graph are semantic supervoxels that capture the spatiotemporal structure of objects and motion cues in the video, while edges between nodes encode spatiotemporal relations and feature similarity. We demonstrate that a shallow network that interleaves graph convolution and graph pooling on this compact representation implements an effective mechanism of relational reasoning yielding strong recognition results on both Charades and Something-Something.

Applying curriculum learning requires both a range of difficulty in data and a method for determining the difficulty of examples. In many tasks, however, satisfying these requirements can be a formidable challenge.In this paper, we contend that histopathology image classification is a compelling use case for curriculum learning. Based on the nature of histopathology images, a range of difficulty inherently exists among examples, and, since medical datasets are often labeled by multiple annotators, annotator agreement can be used as a natural proxy for the difficulty of a given example. Hence, we propose a simple curriculum learning method that trains on progressively-harder images as determined by annotator agreement. We evaluate our hypothesis on the challenging and clinically-important task of colorectal polyp classification. Whereas vanilla training achieves an AUC of 83.7% for this task, a model trained with our proposed curriculum learning approach achieves an AUC of 88.2%, an improvement of 4.5%. Our work aims to inspire researchers to think more creatively and rigorously when choosing contexts for applying curriculum learning.

Developing deep learning models to analyze histology images has been computationally challenging, as the massive size of the images causes excessive strain on all parts of the computing pipeline. This paper proposes a novel deep learning-based methodology for improving the computational efficiency of histology image classification. The proposed approach is robust when used with images that have reduced input resolution, and it can be trained effectively with limited labeled data. Moreover, our approach operates at either the tissue- or slide-level, removing the need for laborious patch-level labeling. Our method uses knowledge distillation to transfer knowledge from a teacher model pre-trained at high resolution to a student model trained on the same images at a considerably lower resolution. Also, to address the lack of large-scale labeled histology image datasets, we perform the knowledge distillation in a self-supervised fashion. We evaluate our approach on three distinct histology image datasets associated with celiac disease, lung adenocarcinoma, and renal cell carcinoma. Our results on these datasets demonstrate that a combination of knowledge distillation and self-supervision allows the student model to approach and, in some cases, surpass the teacher model's classification accuracy while being much more computationally efficient. Additionally, we observe an increase in student classification performance as the size of the unlabeled dataset increases, indicating that there is potential for this method to scale further with additional unlabeled data. Our model outperforms the high-resolution teacher model for celiac disease in accuracy, F1-score, precision, and recall while requiring 4 times fewer computations. For lung adenocarcinoma, our results at 1.25× magnification are within 1.5% of the results for the teacher model at 10× magnification, with a reduction in computational cost by a factor of 64. Our model on renal cell carcinoma at 1.25× magnification performs within 1% of the teacher model at 5× magnification while requiring 16 times fewer computations. Furthermore, our celiac disease outcomes benefit from additional performance scaling with the use of more unlabeled data. In the case of 0.625× magnification, using unlabeled data improves accuracy by 4% over the tissue-level baseline. Therefore, our approach can improve the feasibility of deep learning solutions for digital pathology on standard computational hardware and infrastructures.

Gradient-based approximate inference methods, such as Stein variational gradient descent (SVGD) \cite{liu2016stein}, provide simple and general-purpose inference engines for differentiable continuous distributions. However, existing forms of SVGD can not be directly applied to discrete distributions. In this work, we fill this gap by proposing a simple general-purpose framework that transforms discrete distributions to equivalent piecewise continuous distribution, on which we apply gradient-free Stein variational gradient descent to perform efficient approximate inference. Our empirical results show that our method outperforms traditional algorithms such as Gibbs sampling and discontinuous Hamiltonian Monte Carlo on various challenging benchmarks of discrete graphical models. We demonstrate that our method provides a promising tool for learning ensembles of binarized neural network (BNN), outperforming other widely used ensemble methods on learning binarized AlexNet on CIFAR-10. In addition, such transform can be straightforwardly employed in gradient-free kernelized Stein discrepancy to perform goodness-of-fit (GOF) test on discrete distributions. Our proposed method outperforms existing GOF test methods for intractable discrete distributions.

We introduce a method for simultaneously classifying, segmenting and tracking object instances in a video sequence. Our method, named MaskProp, adapts the popular Mask R-CNN to video by adding a mask propagation branch that propagates frame-level object instance masks from each video frame to all the other frames in a video clip. This allows our system to predict clip-level instance tracks with respect to the object instances segmented in the middle frame of the clip. Clip-level instance tracks generated densely for each frame in the sequence are finally aggregated to produce video-level object instance segmentation and classification. Our experiments demonstrate that our clip-level instance segmentation makes our approach robust to motion blur and object occlusions in video. MaskProp achieves the best reported accuracy on the YouTube-VIS dataset, outperforming the ICCV 2019 video instance segmentation challenge winner despite being much simpler and using orders of magnitude less labeled data (1.3M vs 1B images and 860K vs 14M bounding boxes). The project page is at: https://gberta.github.io/maskprop/.

In the face of the video data deluge, today's expensive clip-level classifiers are increasingly impractical. We propose a framework for efficient action recognition in untrimmed video that uses audio as a preview mechanism to eliminate both short-term and long-term visual redundancies. First, we devise an ImgAud2Vid framework that hallucinates clip-level features by distilling from lighter modalities---a single frame and its accompanying audio---reducing short-term temporal redundancy for efficient clip-level recognition. Second, building on ImgAud2Vid, we further propose ImgAud-Skimming, an attention-based long short-term memory network that iteratively selects useful moments in untrimmed videos, reducing long-term temporal redundancy for efficient video-level recognition. Extensive experiments on four action recognition datasets demonstrate that our method achieves the state-of-the-art in terms of both recognition accuracy and speed.

Motion is a salient cue to recognize actions in video. Modern action recognition models leverage motion information either explicitly by using optical flow as input or implicitly by means of 3D convolutional filters that simultaneously capture appearance and motion information. This paper proposes an alternative approach based on a learnable correlation operator that can be used to establish frame-to-frame matches over convolutional feature maps in the different layers of the network. The proposed architecture enables the fusion of this explicit temporal matching information with traditional appearance cues captured by 2D convolution. Our correlation network compares favorably with widely-used 3D CNNs for video modeling, and achieves competitive results over the prominent two-stream network while being much faster to train. We empirically demonstrate that correlation networks produce strong results on a variety of video datasets, and outperform the state of the art on four popular benchmarks for action recognition: Kinetics, Something-Something, Diving48, and Sports1M.

Visual and audio modalities are highly correlated, yet they contain different information. Their strong correlation makes it possible to predict the semantics of one from the other with good accuracy. Their intrinsic differences make cross-modal prediction a potentially more rewarding pretext task for self-supervised learning of video and audio representations compared to within-modality learning. Based on this intuition, we propose Cross-Modal Deep Clustering (XDC), a novel self-supervised method that leverages unsupervised clustering in one modality (e.g., audio) as a supervisory signal for the other modality (e.g., video). This cross-modal supervision helps XDC utilize the semantic correlation and the differences between the two modalities. Our experiments show that XDC outperforms single-modality clustering and other multi-modal variants. XDC achieves state-of-the-art accuracy among self-supervised methods on multiple video and audio benchmarks. Most importantly, our video model pretrained on large-scale unlabeled data significantly outperforms the same model pretrained with full-supervision on ImageNet and Kinetics for action recognition on HMDB51 and UCF101. To the best of our knowledge, XDC is the first self-supervised learning method that outperforms large-scale fully-supervised pretraining for action recognition on the same architecture.

Many objects in the real world undergo dramatic variations in visual appearance. For example, a tomato may be red or green, sliced or chopped, fresh or fried, liquid or solid. Training a single detector to accurately recognize tomatoes in all these different states is challenging. On the other hand, contextual cues (e.g., the presence of a knife, a cutting board, a strainer or a pan) are often strongly indicative of how the object appears in the scene. Recognizing such contextual cues is useful not only to improve the accuracy of object detection or to determine the state of the object, but also to understand its functional properties and to infer ongoing or upcoming human-object interactions. A fully-supervised approach to recognizing object states and their contexts in the real-world is unfortunately marred by the long-tailed, open-ended distribution of the data, which would effectively require massive amounts of annotations to capture the appearance of objects in all their different forms. Instead of relying on manually-labeled data for this task, we propose a new framework for learning Contextualized OBject Embeddings (COBE) from automatically-transcribed narrations of instructional videos. We leverage the semantic and compositional structure of language by training a visual detector to predict a contextualized word embedding of the object and its associated narration. This enables the learning of an object representation where concepts relate according to a semantic language metric. Our experiments show that our detector learns to predict a rich variety of contextual object information, and that it is highly effective in the settings of few-shot and zero-shot learning.

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 with image based models by 16%. We believe that our approach opens up new approaches for learning spatiotemporal representations from unlabeled video data.

While many action recognition datasets consist of collections of brief, trimmed videos each containing a relevant action, videos in the real-world (e.g., on YouTube) exhibit very different properties: they are often several minutes long, where brief relevant clips are often interleaved with segments of extended duration containing little change. Applying densely an action recognition system to every temporal clip within such videos is prohibitively expensive. Furthermore, as we show in our experiments, this results in suboptimal recognition accuracy as informative predictions from relevant clips are outnumbered by meaningless classification outputs over long uninformative sections of the video. In this paper we introduce a lightweight ``clip-sampling'' model that can efficiently identify the most salient temporal clips within a long video. We demonstrate that the computational cost of action recognition on untrimmed videos can be dramatically reduced by invoking recognition only on these most salient clips. Furthermore, we show that this yields significant gains in recognition accuracy compared to analysis of all clips or randomly selected clips. On Sports1M, our clip sampling scheme elevates the accuracy of an already state-of-the-art action classifier by 7% and reduces by more than 15 times its computational cost.

Group convolution has been shown to offer great computational savings in various 2D convolutional architectures for image classification. It is natural to ask: 1) if group convolution can help to alleviate the high computational cost of video classification networks; 2) what factors matter the most in 3D group convolutional networks; and 3) what are good computation/accuracy trade-offs with 3D group convolutional networks. This paper studies the effects of different design choices in 3D group convolutional networks for video classification. We empirically demonstrate that the amount of channel interactions plays an important role in the accuracy of 3D group convolutional networks. Our experiments suggest two main findings. First, it is a good practice to factorize 3D convolutions by separating channel interactions and spatiotemporal interactions as this leads to improved accuracy and lower computational cost. Second, 3D channel-separated convolutions provide a form of regularization, yielding lower training accuracy but higher test accuracy compared to 3D convolutions. These two empirical findings lead us to design an architecture -- Channel-Separated Convolutional Network (CSN) -- which is simple, efficient, yet accurate. On Sports1M and Kinetics, our CSNs are comparable with or better than the state-of-the-art while being 2-3 times more efficient.

This paper presents a new large-scale dataset for recognition and temporal localization of human actions collected from Web videos. We refer to it as HACS (Human Action Clips and Segments). We leverage consensus and disagreement among visual classifiers to automatically mine candidate short clips from unlabeled videos, which are subsequently validated by human annotators. The resulting dataset is dubbed HACS Clips. Through a separate process we also collect annotations defining action segment boundaries. This resulting dataset is called HACS Segments. Overall, HACS Clips consists of 1.5M annotated clips sampled from 504K untrimmed videos, and HACS Segments contains 139K action segments densely annotated in 50K untrimmed videos spanning 200 action categories. HACS Clips contains more labeled examples than any existing video benchmark. This renders our dataset both a large-scale action recognition benchmark and an excellent source for spatiotemporal feature learning. In our transfer learning experiments on three target datasets, HACS Clips outperforms Kinetics-600, Moments-In-Time and Sports1M as a pretraining source. On HACS Segments, we evaluate state-of-the-art methods of action proposal generation and action localization, and highlight the new challenges posed by our dense temporal annotations.

Capsule networks have been shown to be powerful models for image classification, thanks to their ability to represent and capture viewpoint variations of an object. However, the high computational complexity of capsule networks that stems from the recurrent dynamic routing poses a major drawback making their use for large-scale image classification challenging. In this work, we propose Star-Caps a capsule-based network that exploits a straight-through attentive routing to address the drawbacks of capsule networks. By utilizing attention modules augmented by differentiable binary routers, the proposed mechanism estimates the routing coefficients between capsules without recurrence, as opposed to prior related work. Subsequently, the routers utilize straight-through estimators to make binary decisions to either connect or disconnect the route between capsules, allowing stable and faster performance. The experiments conducted on several image classification datasets, including MNIST, SmallNorb, CIFAR-10, CIFAR-100, and ImageNet show that Star-Caps outperforms the baseline capsule networks.

Modern approaches for multi-person pose estimation in video require large amounts of dense annotations. However, labeling every frame in a video is costly and labor intensive. To reduce the need for dense annotations, we propose a PoseWarper network that leverages training videos with sparse annotations (every k frames) to learn to perform dense temporal pose propagation and estimation. Given a pair of video frames---a labeled Frame A and an unlabeled Frame B---we train our model to predict human pose in Frame A using the features from Frame B by means of deformable convolutions to implicitly learn the pose warping between A and B. We demonstrate that we can leverage our trained PoseWarper for several applications. First, at inference time we can reverse the application direction of our network in order to propagate pose information from manually annotated frames to unlabeled frames. This makes it possible to generate pose annotations for the entire video given only a few manually-labeled frames. Compared to modern label propagation methods based on optical flow, our warping mechanism is much more compact (6M vs 39M parameters), and also more accurate (88.7% mAP vs 83.8% mAP). We also show that we can improve the accuracy of a pose estimator by training it on an augmented dataset obtained by adding our propagated poses to the original manual labels. Lastly, we can use our PoseWarper to aggregate temporal pose information from neighboring frames during inference. This allows us to obtain state-of-the-art pose detection results on PoseTrack2017 and PoseTrack2018 datasets.

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 [17] and video understanding [5]. 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 [1].

The ability to capture temporal information has been critical to the development of video understanding models. While there have been numerous attempts at modeling motion in videos, an explicit analysis of the effect of temporal information for video understanding is still missing. In this work, we aim to bridge this gap and ask the following question: How important is the motion in the video for recognizing the action? To this end, we propose two novel frameworks: (i) class-agnostic temporal generator and (ii) motion-invariant frame selector to reduce/remove motion for an ablation analysis without introducing other artifacts. This isolates the analysis of motion from other aspects of the video. The proposed frameworks provide a much tighter estimate of the effect of motion (from 25% to 6% on UCF101 and 15% to 5% on Kinetics) compared to baselines in our analysis. Our analysis provides critical insights about existing models like C3D, and how it could be made to achieve comparable results with a sparser set of frames.

In this paper we discuss several forms of spatiotemporal convolutions for video analysis and study their effects on action recognition. Our motivation stems from the observation that 2D CNNs applied to individual frames of the video have remained solid performers in action recognition. In this work we empirically demonstrate the accuracy advantages of 3D CNNs over 2D CNNs within the framework of residual learning. Furthermore, we show that factorizing the 3D convolutional filters into separate spatial and temporal components yields significantly gains in accuracy. Our empirical study leads to the design of a new spatiotemporal convolutional block "R(2+1)D" which produces CNNs that achieve results comparable or superior to the state-of-the-art on Sports-1M, Kinetics, UCF101, and HMDB51.

There is a natural correlation between the visual and auditive elements of a video. In this work we leverage this connection to learn general and effective models for both audio and video analysis from self-supervised temporal synchronization. We demonstrate that a calibrated curriculum learning scheme, a careful choice of negative examples, and the use of a contrastive loss are critical ingredients to obtain powerful multi-sensory representations from models optimized to discern temporal synchronization of audio-video pairs. Without further fine-tuning, the resulting audio features achieve performance superior or comparable to the state-of-the-art on established audio classification benchmarks (DCASE2014 and ESC-50). At the same time, our visual subnet provides a very effective initialization to improve the accuracy of video-based action recognition models: compared to learning from scratch, our self-supervised pretraining yields a remarkable gain of +19.9% in action recognition accuracy on UCF101 and a boost of +17.7% on HMDB51.

We introduce an architecture for image categorization that enables the end-to-end learning of separate visual features for the different classes to distinguish. The proposed model consists of a deep CNN shaped like a tree. The stem of the tree includes a sequence of convolutional layers common to all classes. The stem then splits into multiple branches implementing parallel feature extractors, which are ultimately connected to the final classification layer via learned gated connections. These learned gates determine for each individual class the subset of features to use. Such a scheme naturally encourages the learning of a heterogeneous set of specialized features through the separate branches and it allows each class to use the subset of features that are optimal for its recognition. We show the generality of our proposed method by reshaping several popular CNNs from the literature into our proposed architecture. Our experiments on the CIFAR100, CIFAR10, ImageNet, and Synth datasets show that in each case our resulting model yields a substantial improvement in accuracy over the original CNN. Our empirical analysis also suggests that our scheme acts as a form of beneficial regularization improving generalization performance.

Conditional random fields (CRFs) provide a powerful tool for structured prediction, but cast significant challenges in both the learning and inference steps. Approximation techniques are widely used in both steps, which should be considered jointly to guarantee good performance (a.k.a. “inferning"). Perturb-and-MAP models provide a promising alternative to CRFs, but require global combinatorial optimization and hence they are usable only on specific models. In this work, we present a new Local Perturb-and-MAP (locPMAP) framework that replaces the global optimization with a local optimization by exploiting our observed connection between locPMAP and the pseudolikelihood of the original CRF model. We test our approach on three different vision tasks and show that our method achieves consistently improved performance over other approximate inference techniques optimized to a pseudolikelihood objective. Additionally, we demonstrate that we can integrate our method in the fully convolutional network framework to increase our model’s complexity. Finally, our observed connection between locPMAP and the pseudolikelihood leads to a novel perspective for understanding and using pseudolikelihood.

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just 131 additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

We study the design of deep architectures for lossy image compression. We present two architectural recipes in the context of multi-stage progressive encoders and empirically demonstrate their importance on compression performance. Specifically, we show that: 1) predicting the original image data from residuals in a multi-stage progressive architecture facilitates learning and leads to improved performance at approximating the original content and 2) learning to inpaint (from neighboring image pixels) before performing compression reduces the amount of information that must be stored to achieve a high-quality approximation. Incorporating these design choices in a baseline progressive encoder yields an average reduction of over 60% in file size with similar quality compared to the original residual encoder.

The state-of-the-art in semantic segmentation is currently represented by fully convolutional networks (FCNs). However, FCNs use large receptive fields and many pooling layers, both of which cause blurring and low spatial resolution in the deep layers. As a result FCNs tend to produce segmentations that are poorly localized around object boundaries. Prior work has attempted to address this issue in post-processing steps, for example using a color-based CRF on top of the FCN predictions. However, these approaches require additional parameters and low-level features that are difficult to tune and integrate into the original network architecture. Additionally, most CRFs use colorbased pixel affinities, which are not well suited for semantic segmentation and lead to spatially disjoint predictions. To overcome these problems, we introduce a Boundary Neural Field (BNF), which is a global energy model integrating FCN predictions with boundary cues. The boundary information is used to enhance semantic segment coherence and to improve object localization. Specifically, we first show that the convolutional filters of semantic FCNs provide good features for boundary detection. We then employ the predicted boundaries to define pairwise potentials in our energy. Finally, we show that our energy decomposes semantic segmentation into multiple binary problems, which can be relaxed for efficient global optimization. We report extensive experiments demonstrating that minimization of our global boundary-based energy yields results superior to prior globalization methods, both quantitatively as well as qualitatively.

In this paper we propose a method for estimating depth from a single image using a coarse to fine approach. We argue that modeling the fine depth details is easier after a coarse depth map has been computed. We express a global (coarse) depth map of an image as a linear combination of a depth basis learned from training examples. The depth basis captures spatial and statistical regularities and reduces the problem of global depth estimation to the task of predicting the input-specific coefficients in the linear combination. This is formulated as a regression problem from a holistic representation of the image. Crucially, the depth basis and the regression function are coupled and jointly optimized by our learning scheme. We demonstrate that this results in a significant improvement in accuracy compared to direct regression of depth pixel values or approaches learning the depth basis disjointly from the regression function. The global depth estimate is then used as a guidance by a local refinement method that introduces depth details that were not captured at the global level. Experiments on the NYUv2 and KITTI datasets show that our method outperforms the existing state-of-the-art at a considerably lower computational cost for both training and testing.

This paper introduces self-taught object localization, a novel approach that leverages deep convolutional networks trained for whole-image recognition to localize objects in images without additional human supervision, i.e., without using any ground-truth bounding boxes for training. The key idea is to analyze the change in the recognition scores when artificially masking out different regions of the image. The masking out of a region that includes the object typically causes a significant drop in recognition score. This idea is embedded into an agglomerative clustering technique that generates self-taught localization hypotheses. Our object localization scheme outperforms existing proposal methods in both precision and recall for small number of subwindow proposals (e.g., on ILSVRC-2012 it produces a relative gain of 23.4% over the state-of-the-art for top-1 hypothesis). Furthermore, our experiments show that the annotations automatically-generated by our method can be used to train object detectors yielding recognition results remarkably close to those obtained by training on manually-annotated bounding boxes.

Contour detection has been a fundamental component in many image segmentation and object detection systems. Most previous work utilizes low-level features such as texture or saliency to detect contours and then use them as cues for a higher-level task such as object detection. However, we claim that recognizing objects and predicting contours are two mutually related tasks. Contrary to traditional approaches, we show that we can invert the commonly established pipeline: instead of detecting contours with low-level cues for a higher-level recognition task, we exploit object-related features as high-level cues for contour detection.

Most of the current boundary detection systems rely exclusively on low-level features, such as color and texture. However, perception studies suggest that humans employ object-level reasoning when judging if a particular pixel is a boundary. Inspired by this observation, in this work we show how to predict boundaries by exploiting object-level features from a pretrained object-classification network. Our method can be viewed as a "High-for-Low" approach where high-level object features inform the low-level boundary detection process. Our model achieves state-of-the-art performance on an established boundary detection benchmark and it is efficient to run. Additionally, we show that due to the semantic nature of our boundaries we can use them to aid a number of high-level vision tasks. We demonstrate that using our boundaries we improve the performance of state-of-the-art methods on the problems of semantic boundary labeling, semantic segmentation and object proposal generation. We can view this process as a "Low-for-High'" scheme, where low-level boundaries aid high-level vision tasks. Thus, our contributions include a boundary detection system that is accurate, efficient, generalizes well to multiple datasets, and is also shown to improve existing state-of-the-art high-level vision methods on three distinct tasks.

We propose a simple, yet effective approach for spatiotemporal feature learning using deep 3-dimensional convolutional networks (3D ConvNets) trained on a large scale supervised video dataset. Our findings are three-fold: 1) 3D ConvNets are more suitable for spatiotemporal feature learning compared to 2D ConvNets, 2) A homogeneous architecture with small 3x3x3 convolution kernels in all layers is among the best performing architectures for 3D ConvNets, and 3) Our learned features, namely C3D (Convolutional 3D), with a simple linear classifier outperform state-of-the-art methods on 4 different benchmarks and are comparable with current best methods on the other 2 benchmarks. In addition, the features are compact: achieving 52.8% accuracy on UCF101 dataset with only 10 dimensions and also very efficient to compute due to the fast inference of ConvNets. Finally, they are conceptually very simple and easy to train and use.

Traditional Web search engines do not use the images in the HTML pages to find relevant documents for a given query. Instead, they typically operate by computing a measure of agreement between the keywords provided by the user and only the text portion of each page. In this paper we study whether the content of the pictures appearing in a Web page can be used to enrich the semantic description of an HTML document and consequently boost the performance of a keyword-based search engine. We present a Web-scalable system that exploits a pure text-based search engine to find an initial set of candidate documents for a given query. Then, the candidate set is reranked using visual information extracted from the images contained in the pages. The resulting system retains the computational efficiency of traditional text-based search engines with only a small additional storage cost needed to encode the visual information. We test our approach on one of the TREC Million Query Track benchmarks where we show that the exploitation of visual content yields improvement in accuracies for two distinct text-based search engines, including the system with the best reported performance on this benchmark. We further validate our approach by collecting document relevance judgements on our search results using Amazon Mechanical Turk. The results of this experiment confirm the improvement in accuracy produced by our image-based reranker over a pure text-based system.

AutoCaption is a system that helps a smartphone user generate a caption for their photos. It operates by uploading the photo to a cloud service where a number of parallel modules are applied to recognize a variety of entities and relations. The outputs of the modules are combined to generate a large set of candidate captions, which are returned to the phone. The phone client includes a convenient user interface that allows users to select their favorite caption, reorder, add, or delete words to obtain the grammatical style they prefer. The user can also select from multiple candidates returned by the recognition modules.

This paper describes compact image descriptors enabling accurate object categorization with linear classification models, which offer the advantage of being efficient to both train and test. The shared property of our descriptors is the use of classifiers to produce the features of each image. Intuitively, these classifiers evaluate the presence of a set of basis classes inside the image. We first propose to train the basis classifiers as recognizers of a hand-selected set of object classes. We then demonstrate that better accuracy can be achieved by learning the basis classes as "abstract categories" collectively optimized as features for linear classification. Finally, we describe several strategies to aggregate the outputs of basis classifiers evaluated on multiple subwindows of the image in order to handle cases when the photo contains multiple objects and large amounts of clutter. We test our descriptors on challenging benchmarks of object categorization and detection, using a simple linear SVM as classifier. Our results are on par with those achieved by the best systems in these fields but are produced at orders of magnitude lower computational costs and using an image representation that is general and not specifically tuned for a predefined set of test classes.

In this paper we propose a new technique for learning a discriminative codebook for local feature descriptors, specifically designed for scalable landmark classification. The key contribution lies in exploiting the knowledge of correspondences within sets of feature descriptors during code-book learning. Feature correspondences are obtained using structure from motion (SfM) computation on Internet photo collections which serve as the training data. Our codebook is defined by a random forest that is trained to map corresponding feature descriptors into identical codes. Unlike prior forest-based codebook learning methods, we utilize fine-grained descriptor labels and address the challenge of training a forest with an extremely large number of labels. Our codebook is used with various existing feature encoding schemes and also a variant we propose for importance-weighted aggregation of local features. We evaluate our approach on a public dataset of 25 landmarks and our new dataset of 620 landmarks (614K images). Our approach significantly outperforms the state of the art in landmark classification. Furthermore, our method is memory efficient and scalable.

We present CarSafe, a new driver safety app for Android phones that detects and alerts drivers to dangerous driving conditions and behavior. It uses computer vision and machine learning algorithms on the phone to monitor and detect whether the driver is tired or distracted using the front-facing camera while at the same time tracking road conditions using the rear-facing camera. Today's smartphones do not, however, have the capability to process video streams from both the front and rear cameras simultaneously. In response, CarSafe uses acontext-aware algorithm that switches between the two cameras while processing the data in real-time with the goal of minimizing missed events inside (e.g., drowsy driving) and outside of the car (e.g., tailgating). Camera switching means that CarSafe technically has a "blind spot" in the front or rear at any given time. To address this, CarSafe uses other embedded sensors on the phone (i.e., inertial sensors) to generate soft hints regarding potential blind spot dangers. We present the design and implementation of CarSafe and discuss its evaluation using results from a 12-driver field trial. Results from the CarSafe deployment are promising -- CarSafe can infer a common set of dangerous driving behaviors and road conditions with an overall precision and recall of 83% and 75%, respectively. CarSafe is the first dual-camera sensing app for smartphones and represents a new disruptive technology because it provides similar advanced safety features otherwise only found in expensive top-end cars.

We present CarSafe, the first driver safety application that uses dual cameras on smartphones to detect and alert drivers to dangerous driving conditions. CarSafe fuses events detected from cameras and readings from embedded sensors on the phone -- such as the GPS, accelerometer and gyroscope -- to detect and alert the driver of dangerous driving behavior in and outside of the car. Results from a 12-driver field trial show CarSafe can infer five of the most commonly occurring dangerous driving conditions with an overall precision and recall of 83% and 75%, respectively.

Traditional Web search engines do not use the images in the HTML pages to find relevant documents for a given query. Instead, they typically operate by computing a measure of agreement between the keywords provided by the user and only the text portion of each page. In this paper we study whether the content of the pictures appearing in a Web page can be used to enrich the semantic description of an HTML document and consequently boost the performance of a keyword-based search engine. We present a Web-scalable system that exploits a pure text-based search engine to find an initial set of candidate documents for a given query. Then, the candidate set is reranked using semantic information extracted from the images contained in the pages. The resulting system retains the computational efficiency of traditional text-based search engines with only a small additional storage cost needed to encode the visual information. We test our approach on the TREC 2009 Million Query Track, where we show that our use of visual content yields improvement in accuracies for two distinct text-based search engines, including the system with the best reported performance on this benchmark.

In this paper, we present a new approach for establishing correspondences between sparse image features related by an unknown nonrigid mapping and corrupted by clutter and occlusion, such as points extracted from images of different instances of the same object category. We formulate this matching task as an energy minimization problem by defining an elaborate objective function of the appearance and the spatial arrangement of the features. Optimization of this energy is an instance of graph matching, which is in general an NP-hard problem. We describe a novel graph matching optimization technique, which we refer to as dual decomposition (DD), and demonstrate on a variety of examples that this method outperforms existing graph matching algorithms. In the majority of our examples, DD is able to find the global minimum within a minute. The ability to globally optimize the objective allows us to accurately learn the parameters of our matching model from training examples. We show on several matching tasks that our learned model yields results superior to those of state-of-the-art methods.

In this paper we introduce a novel image descriptor enabling accurate object categorization even with linear models. Akin to the popular attribute descriptors, our feature vector comprises the outputs of a set of classifiers evaluated on the image. However, unlike traditional attributes which represent hand-selected object classes and predefined visual properties, our features are learned automatically and correspond to “abstract” categories, which we name meta-classes. Each meta-class is a super-category obtained by grouping a set of object classes such that, collectively, they are easy to distinguish from other sets of categories. By using “learnability” of the meta-classes as criterion for feature generation, we obtain a set of attributes that encode general visual properties shared by multiple object classes and that are effective in describing and recognizing even novel categories, i.e., classes not present in the training set. We demonstrate that simple linear SVMs trained on our meta-class descriptor significantly outperform the best known classifier on the Caltech256 benchmark. We also present results on the 2010 ImageNet Challenge database where our system produces results approaching those of the best systems, but at a much lower computational cost.

We demonstrate CarSafe, a driver safety application for Android phones that fuses information from both front and back cameras and others embedded sensors on the phone to detect and alert drivers to dangerous driving conditions in and outside of the car. In this demonstration, we set up an emulated driving environment to show how CarSafe works.

Driving while being tired or distracted is dangerous. We are developing the CafeSafe app for Android phones, which fuses information from both front and back cameras and others embedded sensors on the phone to detect and alert drivers to dangerous driving conditions in and outside of the car. CarSafe uses computer vision and machine learning algorithms on the phone to monitor and detect whether the driver is tired or distracted using the front camera while at the same time tracking road conditions using the back camera. CarSafe is the first dual-camera application for smart-phones.

In this paper we address the problem of object-class retrieval in large image data sets: given a small set of training examples defining a visual category, the objective is to efficiently retrieve images of the same class from a large database. We propose two contrasting retrieval schemes achieving good accuracy and high efficiency. The first exploits sparse classification models expressed as linear combinations of a small number of features. These sparse models can be efficiently evaluated using inverted file indexing. Furthermore, we introduce a novel ranking procedure that provides a significant speedup over inverted file indexing when the goal is restricted to finding the top-k (i.e., the k highest ranked) images in the data set. We contrast these sparse retrieval models with a second scheme based on approximate ranking using vector quantization. Experimental results show that our algorithms for object-class retrieval can search a 10 million database in just a couple of seconds and produce categorization accuracy comparable to the best known class-recognition systems.

We introduce PiCoDes: a very compact image descriptor which nevertheless allows high performance on object category recognition. In particular, we address novel-category recognition: the task of defining indexing structures and image representations which enable a large collection of images to be searched for an object category that was not known when the index was built. Instead, the training images defining the category are supplied at query time. We explicitly learn descriptors of a given length (from as small as 16 bytes per image) which have good object-recognition performance. In contrast to previous work in the domain of object recognition, we do not choose an arbitrary intermediate representation, but explicitly learn short codes. In contrast to previous approaches to learn compact codes, we optimize explicitly for (an upper bound on) classification performance. Optimization directly for binary features is difficult and nonconvex, but we present an alternation scheme and convex upper bound which demonstrate excellent performance in practice. PiCoDes of 256 bytes match the accuracy of the current best known classifier for the Caltech256 benchmark, but they decrease the database storage size by a factor of 100 and speed-up the training and testing of novel classes by orders of magnitude.

It has been shown that the 3D shape of a deformable surface in an image can be recovered by establishing correspondences between that image and a reference one in which the shape is known. These matches can then be used to set-up a convex optimization problem in terms of the shape parameters, which is easily solved. However, in many cases, the correspondences are hard to establish reliably. In this paper, we show that we can solve simultaneously for both 3D shape and correspondences, thereby using 3D shape constraints to guide the image matching and increasing robustness, for example when the textures are repetitive. This involves solving a mixed integer quadratic problem. While optimizing this problem is NP-hard in general, we show that its solution can nevertheless be approximated effectively by a branch-and-bound algorithm.

Most current image categorization methods require large collections of manually annotated training examples to learn accurate visual recognition models. The time-consuming human labeling effort effectively limits these approaches to recognition problems involving a small number of different object classes. In order to address this shortcoming, in recent years several authors have proposed to learn object classifiers from weakly-labeled Internet images, such as photos retrieved by keyword-based image search engines. While this strategy eliminates the need for human supervision, the recognition accuracies of these methods are considerably lower than those obtained with fully-supervised approaches, because of the noisy nature of the labels associated to Web data. In this paper we investigate and compare methods that learn image classifiers by combining very few manually annotated examples (e.g., 1-10 images per class) and a large number of weakly-labeled Web photos retrieved using keyword-based image search. We cast this as a domain adaptation problem: given a few strongly-labeled examples in a target domain (the manually annotated examples) and many source domain examples (the weakly-labeled Web photos), learn classifiers yielding small generalization error on the target domain. Our experiments demonstrate that, for the same number of strongly-labeled examples, our domain adaptation approach produces significant recognition rate improvements over the best published results (e.g., 65% better when using 5 labeled training examples per class) and that our classifiers are one order of magnitude faster to learn and to evaluate than the best competing method, despite our use of large weakly-labeled data sets.

We describe an algorithm for similar-image search which is designed to be efficient for extremely large collections of images. For each query, a small response set is selected by a fast prefilter, after which a more accurate ranker may be applied to each image in the response set. We consider a class of prefilters comprising disjunctions of conjunctions (“ORs of ANDs”) of Boolean features. AND filters can be implemented efficiently using skipped inverted files, a key component of Web-scale text search engines. These structures permit search in time proportional to the response set size. The prefilters are learned from training examples, and refined at query time to produce an approximately bounded response set. We cast prefiltering as an optimization problem: for each test query, select the OR-of-AND filter which maximizes training-set recall for an adjustable bound on response set size. This may be efficiently implemented by selecting from a large pool of candidate conjunctions of Boolean features using a linear program relaxation. Tests on object class recognition show that this relatively simple filter is nevertheless powerful enough to capture some semantic information.

Visual categorization problems, such as object classification or action recognition, are increasingly often approached using a detection strategy: a classifier function is first applied to candidate subwindows of the image or the video, and then the maximum classifier score is used for class decision. Traditionally, the subwindow classifiers are trained on a large collection of examples manually annotated with masks or bounding boxes. The reliance on time-consuming human labeling effectively limits the application of these methods to problems involving very few categories. Furthermore, the human selection of the masks introduces arbitrary biases (e.g. in terms of window size and location) which may be suboptimal for classification. In this paper we propose a novel method for learning a discriminative subwindow classifier from examples annotated with binary labels indicating the presence of an object or action of interest, but not its location. During training, our approach simultaneously localizes the instances of the positive class and learns a subwindow SVM to recognize them. We extend our method to classification of time series by presenting an algorithm that localizes the most discriminative set of temporal segments in the signal. We evaluate our approach on several datasets for object and action recognition and show that it achieves results similar and in many cases superior to those obtained with full supervision.

This paper describes an unsupervised learning technique for modeling human locomotion styles, such as distinct related activities (e.g. running and striding) or variations of the same motion performed by different subjects. Modeling motion styles requires identifying the common structure in the motions and detecting stylespecific characteristics. We propose an algorithm that learns a hierarchical model of styles from unlabeled motion capture data by exploiting the cyclic property of human locomotion. We assume that sequences with the same style contain locomotion cycles generated by noisy, temporally warped versions of a single latent cycle. We model these style-specific latent cycles as random variables drawn from a common ”parent” cycle distribution, representing the structure shared by all motions. Given these hierarchical priors, the algorithm learns, in a completely unsupervised fashion, temporally aligned latent cycle distributions, each modeling a specific locomotion style, and computes for each example the style label posterior distribution, the segmentation into cycles, and the temporal warping with respect to the latent cycles. We demonstrate the flexibility of the model on several application problems such as style clustering, animation, style blending, and filling in of missing data. 1.

This paper describes methods for recovering time-varying shape and motion of non-rigid 3D objects from uncalibrated 2D point tracks. For example, given a video recording of a talking person, we would like to estimate the 3D shape of the face at each instant, and learn a model of facial deformation. Time-varying shape is modeled as a rigid transformation combined with a non-rigid deformation. Reconstruction is ill-posed if arbitrary deformations are allowed, and thus additional assumptions about deformations are required. We first suggest restricting shapes to lie within a low-dimensional subspace, and describe estimation algorithms. However, this restriction alone is insufficient to constrain reconstruction. To address these problems, we propose a reconstruction method using a Probabilistic Principal Components Analysis (PPCA) shape model, and an estimation algorithm that simultaneously estimates 3D shape and motion for each instant, learns the PPCA model parameters, and robustly fills-in missing data points. We then extend the model to model temporal dynamics in object shape, allowing the algorithm to robustly handle severe cases of missing data.

This paper presents an algorithm for synthesis of human motion in specified styles. We use a theory of movement observation (Laban Movement Analysis) to describe movement styles as points in a multi-dimensional perceptual space. We cast the task of learning to synthesize desired movement styles as a regression problem: sequences generated via space-time interpolation of motion capture data are used to learn a nonlinear mapping between animation parameters and movement styles in perceptual space. We demonstrate that the learned model can apply a variety of motion styles to pre-recorded motion sequences and it can extrapolate styles not originally included in the training data.

Metric learning has been shown to significantly improve the accuracy of k-nearest neighbor (kNN) classification. In problems involving thousands of features, dis- tance learning algorithms cannot be used due to overfitting and high computa- tional complexity. In such cases, previous work has relied on a two-step solution: first apply dimensionality reduction methods to the data, and then learn a met- ric in the resulting low-dimensional subspace. In this paper we show that better classification performance can be achieved by unifying the objectives of dimen- sionality reduction and metric learning. We propose a method that solves for the low-dimensional projection of the inputs, which minimizes a metric objective aimed at separating points in different classes by a large margin. This projection is defined by a significantly smaller number of parameters than metrics learned in input space, and thus our optimization reduces the risks of overfitting. Theory and results are presented for both a linear as well as a kernelized version of the algorithm. Overall, we achieve classification rates similar, and in several cases superior, to those of support vector machines.

This paper presents an algorithm for learning the time-varying shape of a non-rigid 3D object from uncalibrated 2D tracking data. We model shape motion as a rigid component (rotation and translation) combined with a non-rigid deformation. Reconstruction is ill-posed if arbitrary deforma- tions are allowed. We constrain the problem by assuming that the object shape at each time instant is drawn from a Gaussian distribution. Based on this assumption, the algorithm simultaneously estimates 3D shape and motion for each time frame, learns the parameters of the Gaussian, and robustly fills-in missing data points. We then extend the algorithm to model temporal smoothness in object shape, thus allowing it to handle severe cases of missing data.

This paper presents a novel solution for flow-based tracking and 3D reconstruction of deforming objects in monocular image sequences. A non-rigid 3D object undergoing rotation and deformation can be effectively approximated using a linear combination of 3D basis shapes. This puts a bound on the rank of the tracking matrix. The rank constraint is used to achieve robust and precise low-level optical flow estimation without prior knowledge of the 3D shape of the object. The bound on the rank is also exploited to handle occlusion at the tracking level leading to the possibility of recovering the complete trajectories of occluded/disoccluded points. Following the same low-rank principle, the resulting flow matrix can be factored to get the 3D pose, configuration coefficients, and 3D basis shapes. The flow matrix is factored in an iterative manner, looping between solving for pose, configuration, and basis shapes. The flow-based tracking is applied to several video sequences and provides the input to the 3D non-rigid reconstruction task. Additional results on synthetic data and comparisons to ground truth complete the experiments.