Search

Understanding what deep network models capture in their learned representations is a fundamental challenge in computer vision. We present a new methodology to understanding such vision models, the Visual Concept Connectome (VCC), which discovers human interpretable concepts and their interlayer connections in a fully unsupervised manner. Our approach simultaneously reveals fine-grained concepts at a layer, connection weightings across all layers and is amendable to global analysis of network structure (e.g., branching pattern of hierarchical concept assemblies). Previous work yielded ways to extract interpretable concepts from single layers and examine their impact on classification, but did not afford multilayer concept analysis across an entire network architecture. Quantitative and qualitative empirical results show the effectiveness of VCCs in the domain of image classification. Also, we leverage VCCs for the application of failure mode debugging to reveal where mistakes arise in deep networks., Comment: CVPR 2024 (Highlight)

Concerns for the privacy of individuals captured in public imagery have led to privacy-preserving action recognition. Existing approaches often suffer from issues arising through obfuscation being applied globally and a lack of interpretability. Global obfuscation hides privacy sensitive regions, but also contextual regions important for action recognition. Lack of interpretability erodes trust in these new technologies. We highlight the limitations of current paradigms and propose a solution: Human selected privacy templates that yield interpretability by design, an obfuscation scheme that selectively hides attributes and also induces temporal consistency, which is important in action recognition. Our approach is architecture agnostic and directly modifies input imagery, while existing approaches generally require architecture training. Our approach offers more flexibility, as no retraining is required, and outperforms alternatives on three widely used datasets.

Video segmentation encompasses a wide range of categories of problem formulation, e.g., object, scene, actor-action and multimodal video segmentation, for delineating task-specific scene components with pixel-level masks. Recently, approaches in this research area shifted from concentrating on ConvNet-based to transformer-based models. In addition, various interpretability approaches have appeared for transformer models and video temporal dynamics, motivated by the growing interest in basic scientific understanding, model diagnostics and societal implications of real-world deployment. Previous surveys mainly focused on ConvNet models on a subset of video segmentation tasks or transformers for classification tasks. Moreover, component-wise discussion of transformer-based video segmentation models has not yet received due focus. In addition, previous reviews of interpretability methods focused on transformers for classification, while analysis of video temporal dynamics modelling capabilities of video models received less attention. In this survey, we address the above with a thorough discussion of various categories of video segmentation, a component-wise discussion of the state-of-the-art transformer-based models, and a review of related interpretability methods. We first present an introduction to the different video segmentation task categories, their objectives, specific challenges and benchmark datasets. Next, we provide a component-wise review of recent transformer-based models and document the state of the art on different video segmentation tasks. Subsequently, we discuss post-hoc and ante-hoc interpretability methods for transformer models and interpretability methods for understanding the role of the temporal dimension in video models. Finally, we conclude our discussion with future research directions.

Few-shot video segmentation is the task of delineating a specific novel class in a query video using few labelled support images. Typical approaches compare support and query features while limiting comparisons to a single feature layer and thereby ignore potentially valuable information. We present a meta-learned Multiscale Memory Comparator (MMC) for few-shot video segmentation that combines information across scales within a transformer decoder. Typical multiscale transformer decoders for segmentation tasks learn a compressed representation, their queries, through information exchange across scales. Unlike previous work, we instead preserve the detailed feature maps during across scale information exchange via a multiscale memory transformer decoding to reduce confusion between the background and novel class. Integral to the approach, we investigate multiple forms of information exchange across scales in different tasks and provide insights with empirical evidence on which to use in each task. The overall comparisons among query and support features benefit from both rich semantics and precise localization. We demonstrate our approach primarily on few-shot video object segmentation and an adapted version on the fully supervised counterpart. In all cases, our approach outperforms the baseline and yields state-of-the-art performance. Our code is publicly available at https://github.com/MSiam/MMC-MultiscaleMemory.

Instructional videos are an important resource to learn procedural tasks from human demonstrations. However, the instruction steps in such videos are typically short and sparse, with most of the video being irrelevant to the procedure. This motivates the need to temporally localize the instruction steps in such videos, i.e. the task called key-step localization. Traditional methods for key-step localization require video-level human annotations and thus do not scale to large datasets. In this work, we tackle the problem with no human supervision and introduce StepFormer, a self-supervised model that discovers and localizes instruction steps in a video. StepFormer is a transformer decoder that attends to the video with learnable queries, and produces a sequence of slots capturing the key-steps in the video. We train our system on a large dataset of instructional videos, using their automatically-generated subtitles as the only source of supervision. In particular, we supervise our system with a sequence of text narrations using an order-aware loss function that filters out irrelevant phrases. We show that our model outperforms all previous unsupervised and weakly-supervised approaches on step detection and localization by a large margin on three challenging benchmarks. Moreover, our model demonstrates an emergent property to solve zero-shot multi-step localization and outperforms all relevant baselines at this task., Comment: CVPR'23

In this paper, we present an end-to-end trainable unified multiscale encoder-decoder transformer that is focused on dense prediction tasks in video. The presented Multiscale Encoder-Decoder Video Transformer (MED-VT) uses multiscale representation throughout and employs an optional input beyond video (e.g., audio), when available, for multimodal processing (MED-VT++). Multiscale representation at both encoder and decoder yields three key benefits: (i) implicit extraction of spatiotemporal features at different levels of abstraction for capturing dynamics without reliance on input optical flow, (ii) temporal consistency at encoding and (iii) coarse-to-fine detection for high-level (e.g., object) semantics to guide precise localization at decoding. Moreover, we present a transductive learning scheme through many-to-many label propagation to provide temporally consistent video predictions. We showcase MED-VT/MED-VT++ on three unimodal video segmentation tasks (Automatic Video Object Segmentation (AVOS), actor-action segmentation and Video Semantic Segmentation (VSS)) as well as a multimodal segmentation task (Audio-Visual Segmentation (AVS)). Results show that the proposed architecture outperforms alternative state-of-the-art approaches on multiple benchmarks using only video (and optional audio) as input, without reliance on optical flow. Finally, to document details of the model's internal learned representations, we present a detailed interpretability study, encompassing both quantitative and qualitative analyses., Comment: Extension of CVPR'23 paper for journal submission

There is limited understanding of the information captured by deep spatiotemporal models in their intermediate representations. For example, while evidence suggests that action recognition algorithms are heavily influenced by visual appearance in single frames, no quantitative methodology exists for evaluating such static bias in the latent representation compared to bias toward dynamics. We tackle this challenge by proposing an approach for quantifying the static and dynamic biases of any spatiotemporal model, and apply our approach to three tasks, action recognition, automatic video object segmentation (AVOS) and video instance segmentation (VIS). Our key findings are: (i) Most examined models are biased toward static information. (ii) Some datasets that are assumed to be biased toward dynamics are actually biased toward static information. (iii) Individual channels in an architecture can be biased toward static, dynamic or a combination of the two. (iv) Most models converge to their culminating biases in the first half of training. We then explore how these biases affect performance on dynamically biased datasets. For action recognition, we propose StaticDropout, a semantically guided dropout that debiases a model from static information toward dynamics. For AVOS, we design a better combination of fusion and cross connection layers compared with previous architectures., Comment: TPAMI 2024. arXiv admin note: substantial text overlap with arXiv:2206.02846

This paper investigates the modeling of automated machine description on sports video, which has seen much progress recently. Nevertheless, state-of-the-art approaches fall quite short of capturing how human experts analyze sports scenes. There are several major reasons: (1) The used dataset is collected from non-official providers, which naturally creates a gap between models trained on those datasets and real-world applications; (2) previously proposed methods require extensive annotation efforts (i.e., player and ball segmentation at pixel level) on localizing useful visual features to yield acceptable results; (3) very few public datasets are available. In this paper, we propose a novel large-scale NBA dataset for Sports Video Analysis (NSVA) with a focus on captioning, to address the above challenges. We also design a unified approach to process raw videos into a stack of meaningful features with minimum labelling efforts, showing that cross modeling on such features using a transformer architecture leads to strong performance. In addition, we demonstrate the broad application of NSVA by addressing two additional tasks, namely fine-grained sports action recognition and salient player identification. Code and dataset are available at https://github.com/jackwu502/NSVA.

Intuition might suggest that motion and dynamic information are key to video-based action recognition. In contrast, there is evidence that state-of-the-art deep-learning video understanding architectures are biased toward static information available in single frames. Presently, a methodology and corresponding dataset to isolate the effects of dynamic information in video are missing. Their absence makes it difficult to understand how well contemporary architectures capitalize on dynamic vs. static information. We respond with a novel Appearance Free Dataset (AFD) for action recognition. AFD is devoid of static information relevant to action recognition in a single frame. Modeling of the dynamics is necessary for solving the task, as the action is only apparent through consideration of the temporal dimension. We evaluated 11 contemporary action recognition architectures on AFD as well as its related RGB video. Our results show a notable decrease in performance for all architectures on AFD compared to RGB. We also conducted a complimentary study with humans that shows their recognition accuracy on AFD and RGB is very similar and much better than the evaluated architectures on AFD. Our results motivate a novel architecture that revives explicit recovery of optical flow, within a contemporary design for best performance on AFD and RGB.

Deep spatiotemporal models are used in a variety of computer vision tasks, such as action recognition and video object segmentation. Currently, there is a limited understanding of what information is captured by these models in their intermediate representations. For example, while it has been observed that action recognition algorithms are heavily influenced by visual appearance in single static frames, there is no quantitative methodology for evaluating such static bias in the latent representation compared to bias toward dynamic information (e.g. motion). We tackle this challenge by proposing a novel approach for quantifying the static and dynamic biases of any spatiotemporal model. To show the efficacy of our approach, we analyse two widely studied tasks, action recognition and video object segmentation. Our key findings are threefold: (i) Most examined spatiotemporal models are biased toward static information; although, certain two-stream architectures with cross-connections show a better balance between the static and dynamic information captured. (ii) Some datasets that are commonly assumed to be biased toward dynamics are actually biased toward static information. (iii) Individual units (channels) in an architecture can be biased toward static, dynamic or a combination of the two., Comment: CVPR 2022

In this paper, we study the problem of procedure planning in instructional videos. Here, an agent must produce a plausible sequence of actions that can transform the environment from a given start to a desired goal state. When learning procedure planning from instructional videos, most recent work leverages intermediate visual observations as supervision, which requires expensive annotation efforts to localize precisely all the instructional steps in training videos. In contrast, we remove the need for expensive temporal video annotations and propose a weakly supervised approach by learning from natural language instructions. Our model is based on a transformer equipped with a memory module, which maps the start and goal observations to a sequence of plausible actions. Furthermore, we augment our model with a probabilistic generative module to capture the uncertainty inherent to procedure planning, an aspect largely overlooked by previous work. We evaluate our model on three datasets and show our weaklysupervised approach outperforms previous fully supervised state-of-the-art models on multiple metrics., Comment: Accepted as an oral paper at CVPR 2022

Few-shot video object segmentation (FS-VOS) aims at segmenting video frames using a few labelled examples of classes not seen during initial training. In this paper, we present a simple but effective temporal transductive inference (TTI) approach that leverages temporal consistency in the unlabelled video frames during few-shot inference. Key to our approach is the use of both global and local temporal constraints. The objective of the global constraint is to learn consistent linear classifiers for novel classes across the image sequence, whereas the local constraint enforces the proportion of foreground/background regions in each frame to be coherent across a local temporal window. These constraints act as spatiotemporal regularizers during the transductive inference to increase temporal coherence and reduce overfitting on the few-shot support set. Empirically, our model outperforms state-of-the-art meta-learning approaches in terms of mean intersection over union on YouTube-VIS by 2.8%. In addition, we introduce improved benchmarks that are exhaustively labelled (i.e. all object occurrences are labelled, unlike the currently available), and present a more realistic evaluation paradigm that targets data distribution shift between training and testing sets. Our empirical results and in-depth analysis confirm the added benefits of the proposed spatiotemporal regularizers to improve temporal coherence and overcome certain overfitting scenarios., Comment: IJCV submission under review

Video predictive understanding encompasses a wide range of efforts that are concerned with the anticipation of the unobserved future from the current as well as historical video observations. Action prediction is a major sub-area of video predictive understanding and is the focus of this review. This sub-area has two major subdivisions: early action recognition and future action prediction. Early action recognition is concerned with recognizing an ongoing action as soon as possible. Future action prediction is concerned with the anticipation of actions that follow those previously observed. In either case, the \textbf{\textit{causal}} relationship between the past, current, and potential future information is the main focus. Various mathematical tools such as Markov Chains, Gaussian Processes, Auto-Regressive modeling, and Bayesian recursive filtering are widely adopted jointly with computer vision techniques for these two tasks. However, these approaches face challenges such as the curse of dimensionality, poor generalization, and constraints from domain-specific knowledge. Recently, structures that rely on deep convolutional neural networks and recurrent neural networks have been extensively proposed for improving the performance of existing vision tasks, in general, and action prediction tasks, in particular. However, they have their own shortcomings, \eg reliance on massive training data and lack of strong theoretical underpinnings. In this survey, we start by introducing the major sub-areas of the broad area of video predictive understanding, which recently have received intensive attention and proven to have practical value. Next, a thorough review of various early action recognition and future action prediction algorithms are provided with suitably organized divisions. Finally, we conclude our discussion with future research directions.

Early action recognition (action prediction) from limited preliminary observations plays a critical role for streaming vision systems that demand real-time inference, as video actions often possess elongated temporal spans which cause undesired latency. In this study, we address action prediction by investigating how action patterns evolve over time in a spatial feature space. There are three key components to our system. First, we work with intermediate-layer ConvNet features, which allow for abstraction from raw data, while retaining spatial layout. Second, instead of propagating features per se, we propagate their residuals across time, which allows for a compact representation that reduces redundancy. Third, we employ a Kalman filter to combat error build-up and unify across prediction start times. Extensive experimental results on multiple benchmarks show that our approach leads to competitive performance in action prediction. Notably, we investigate the learned components of our system to shed light on their otherwise opaque natures in two ways. First, we document that our learned feature propagation module works as a spatial shifting mechanism under convolution to propagate current observations into the future. Thus, it captures flow-based image motion information. Second, the learned Kalman filter adaptively updates prior estimation to aid the sequence learning process.

Animals locomote for various reasons: to search for food, find suitable habitat, pursue prey, escape from predators, or seek a mate. The grand scale of biodiversity contributes to the great locomotory design and mode diversity. Various creatures make use of legs, wings, fins and other means to move through the world. In this report, we refer to the locomotion of general biological species as biolocomotion. We present a computational approach to detect biolocomotion in unprocessed video. Significantly, the motion exhibited by the body parts of a biological entity to navigate through an environment can be modeled by a combination of an overall positional advance with an overlaid asymmetric oscillatory pattern, a distinctive signature that tends to be absent in non-biological objects in locomotion. We exploit this key trait of positional advance with asymmetric oscillation along with differences in an object's common motion (extrinsic motion) and localized motion of its parts (intrinsic motion) to detect biolocomotion. An algorithm is developed to measure the presence of these traits in tracked objects to determine if they correspond to a biological entity in locomotion. An alternative algorithm, based on generic features combined with learning is assembled out of components from allied areas of investigation, also is presented as a basis of comparison. A novel biolocomotion dataset encompassing a wide range of moving biological and non-biological objects in natural settings is provided. Also, biolocomotion annotations to an extant camouflage animals dataset are provided. Quantitative results indicate that the proposed algorithm considerably outperforms the alternative approach, supporting the hypothesis that biolocomotion can be detected reliably based on its distinct signature of positional advance with asymmetric oscillation and extrinsic/intrinsic motion dissimilarity.

It has been repeatedly observed that convolutional architectures when applied to image understanding tasks learn oriented bandpass filters. A standard explanation of this result is that these filters reflect the structure of the images that they have been exposed to during training: Natural images typically are locally composed of oriented contours at various scales and oriented bandpass filters are matched to such structure. We offer an alternative explanation based not on the structure of images, but rather on the structure of convolutional architectures. In particular, complex exponentials are the eigenfunctions of convolution. These eigenfunctions are defined globally; however, convolutional architectures operate locally. To enforce locality, one can apply a windowing function to the eigenfunctions, which leads to oriented bandpass filters as the natural operators to be learned with convolutional architectures. From a representational point of view, these filters allow for a local systematic way to characterize and operate on an image or other signal. We offer empirical support for the hypothesis that convolutional networks learn such filters at all of their convolutional layers. While previous research has shown evidence of filters having oriented bandpass characteristics at early layers, ours appears to be the first study to document the predominance of such filter characteristics at all layers. Previous studies have missed this observation because they have concentrated on the cumulative compositional effects of filtering across layers, while we examine the filter characteristics that are present at each layer.

While convolutional neural nets (CNNs) have achieved remarkable performance for a wide range of inverse imaging applications, the filter coefficients are computed in a purely data-driven manner and are not explainable. Inspired by an analytically derived CNN by Hadji et al., in this paper we construct a new layered graph neural net (GNN) using GraphBio as our graph filter. Unlike convolutional filters in previous GNNs, our employed GraphBio is analytically defined and requires no training, and we optimize the end-to-end system only via learning of appropriate graph topology at each layer. In signal filtering terms, it means that our linear graph filter at each layer is always intrepretable as low-pass with known biorthogonal conditions, while the graph spectrum itself is optimized via data training. As an example application, we show that our analytical GNN achieves image denoising performance comparable to a state-of-the-art CNN-based scheme when the training and testing data share the same statistics, and when they differ, our analytical GNN outperforms it by more than 1dB in PSNR., Comment: Image denoising, deep learning, analytical graph filter

This document will review the most prominent proposals using multilayer convolutional architectures. Importantly, the various components of a typical convolutional network will be discussed through a review of different approaches that base their design decisions on biological findings and/or sound theoretical bases. In addition, the different attempts at understanding ConvNets via visualizations and empirical studies will be reviewed. The ultimate goal is to shed light on the role of each layer of processing involved in a ConvNet architecture, distill what we currently understand about ConvNets and highlight critical open problems.

As the success of deep models has led to their deployment in all areas of computer vision, it is increasingly important to understand how these representations work and what they are capturing. In this paper, we shed light on deep spatiotemporal representations by visualizing what two-stream models have learned in order to recognize actions in video. We show that local detectors for appearance and motion objects arise to form distributed representations for recognizing human actions. Key observations include the following. First, cross-stream fusion enables the learning of true spatiotemporal features rather than simply separate appearance and motion features. Second, the networks can learn local representations that are highly class specific, but also generic representations that can serve a range of classes. Third, throughout the hierarchy of the network, features become more abstract and show increasing invariance to aspects of the data that are unimportant to desired distinctions (e.g. motion patterns across various speeds). Fourth, visualizations can be used not only to shed light on learned representations, but also to reveal idiosyncracies of training data and to explain failure cases of the system., Comment: This document is best viewed in Adobe Reader where figures play on click. Supplementary material can be downloaded at http://feichtenhofer.github.io/action_vis.pdf

This paper presents a novel hierarchical spatiotemporal orientation representation for spacetime image analysis. It is designed to combine the benefits of the multilayer architecture of ConvNets and a more controlled approach to spacetime analysis. A distinguishing aspect of the approach is that unlike most contemporary convolutional networks no learning is involved; rather, all design decisions are specified analytically with theoretical motivations. This approach makes it possible to understand what information is being extracted at each stage and layer of processing as well as to minimize heuristic choices in design. Another key aspect of the network is its recurrent nature, whereby the output of each layer of processing feeds back to the input. To keep the network size manageable across layers, a novel cross-channel feature pooling is proposed. The multilayer architecture that results systematically reveals hierarchical image structure in terms of multiscale, multiorientation properties of visual spacetime. To illustrate its utility, the network has been applied to the task of dynamic texture recognition. Empirical evaluation on multiple standard datasets shows that it sets a new state-of-the-art., Comment: accepted at ICCV 2017

Two-stream Convolutional Networks (ConvNets) have shown strong performance for human action recognition in videos. Recently, Residual Networks (ResNets) have arisen as a new technique to train extremely deep architectures. In this paper, we introduce spatiotemporal ResNets as a combination of these two approaches. Our novel architecture generalizes ResNets for the spatiotemporal domain by introducing residual connections in two ways. First, we inject residual connections between the appearance and motion pathways of a two-stream architecture to allow spatiotemporal interaction between the two streams. Second, we transform pretrained image ConvNets into spatiotemporal networks by equipping these with learnable convolutional filters that are initialized as temporal residual connections and operate on adjacent feature maps in time. This approach slowly increases the spatiotemporal receptive field as the depth of the model increases and naturally integrates image ConvNet design principles. The whole model is trained end-to-end to allow hierarchical learning of complex spatiotemporal features. We evaluate our novel spatiotemporal ResNet using two widely used action recognition benchmarks where it exceeds the previous state-of-the-art., Comment: NIPS 2016

In computer vision, action recognition refers to the act of classifying an action that is present in a given video and action detection involves locating actions of interest in space and/or time. Videos, which contain photometric information (e.g. RGB, intensity values) in a lattice structure, contain information that can assist in identifying the action that has been imaged. The process of action recognition and detection often begins with extracting useful features and encoding them to ensure that the features are specific to serve the task of action recognition and detection. Encoded features are then processed through a classifier to identify the action class and their spatial and/or temporal locations. In this report, a thorough review of various action recognition and detection algorithms in computer vision is provided by analyzing the two-step process of a typical action recognition and detection algorithm: (i) extraction and encoding of features, and (ii) classifying features into action classes. In efforts to ensure that computer vision-based algorithms reach the capabilities that humans have of identifying actions irrespective of various nuisance variables that may be present within the field of view, the state-of-the-art methods are reviewed and some remaining problems are addressed in the final chapter.