The widespread adoption of deep neural networks in machine learning calls for an objective quantification of esoteric trust. In this paper we propose GradTrust, a classification trust measure for large-scale neural networks at inference. The proposed method utilizes variance of counterfactual gradients, i.e. the required changes in the network parameters if the label were different. We show that GradTrust is superior to existing techniques for detecting misprediction rates on $50000$ images from ImageNet validation dataset. Depending on the network, GradTrust detects images where either the ground truth is incorrect or ambiguous, or the classes are co-occurring. We extend GradTrust to Video Action Recognition on Kinetics-400 dataset. We showcase results on $14$ architectures pretrained on ImageNet and $5$ architectures pretrained on Kinetics-400. We observe the following: (i) simple methodologies like negative log likelihood and margin classifiers outperform state-of-the-art uncertainty and out-of-distribution detection techniques for misprediction rates, and (ii) the proposed GradTrust is in the Top-2 performing methods on $37$ of the considered $38$ experimental modalities. The code is available at: https://github.com/olivesgatech/GradTrust
Adversarial machine learning attacks on video action recognition models is a growing research area and many effective attacks were introduced in recent years. These attacks show that action recognition models can be breached in many ways. Hence using these models in practice raises significant security concerns. However, there are very few works which focus on defending against or detecting attacks. In this work, we propose a novel universal detection method which is compatible with any action recognition model. In our extensive experiments, we show that our method consistently detects various attacks against different target models with high true positive rates while satisfying very low false positive rates. Tested against four state-of-the-art attacks targeting four action recognition models, the proposed detector achieves an average AUC of 0.911 over 16 test cases while the best performance achieved by the existing detectors is 0.645 average AUC. This 41.2% improvement is enabled by the robustness of the proposed detector to varying attack methods and target models. The lowest AUC achieved by our detector across the 16 test cases is 0.837 while the competing detector's performance drops as low as 0.211. We also show that the proposed detector is robust to varying attack strengths. In addition, we analyze our method's real-time performance with different hardware setups to demonstrate its potential as a practical defense mechanism.
We propose a novel framework for video understanding, called Temporally Contextualized CLIP (TC-CLIP), which leverages essential temporal information through global interactions in a spatio-temporal domain within a video. To be specific, we introduce Temporal Contextualization (TC), a layer-wise temporal information infusion mechanism for videos, which 1) extracts core information from each frame, 2) connects relevant information across frames for the summarization into context tokens, and 3) leverages the context tokens for feature encoding. Furthermore, the Video-conditional Prompting (VP) module processes context tokens to generate informative prompts in the text modality. Extensive experiments in zero-shot, few-shot, base-to-novel, and fully-supervised action recognition validate the effectiveness of our model. Ablation studies for TC and VP support our design choices. Our project page with the source code is available at https://github.com/naver-ai/tc-clip
Zero-Shot Temporal Action Localization (ZS-TAL) seeks to identify and locate actions in untrimmed videos unseen during training. Existing ZS-TAL methods involve fine-tuning a model on a large amount of annotated training data. While effective, training-based ZS-TAL approaches assume the availability of labeled data for supervised learning, which can be impractical in some applications. Furthermore, the training process naturally induces a domain bias into the learned model, which may adversely affect the model's generalization ability to arbitrary videos. These considerations prompt us to approach the ZS-TAL problem from a radically novel perspective, relaxing the requirement for training data. To this aim, we introduce a novel method that performs Test-Time adaptation for Temporal Action Localization (T3AL). In a nutshell, T3AL adapts a pre-trained Vision and Language Model (VLM). T3AL operates in three steps. First, a video-level pseudo-label of the action category is computed by aggregating information from the entire video. Then, action localization is performed adopting a novel procedure inspired by self-supervised learning. Finally, frame-level textual descriptions extracted with a state-of-the-art captioning model are employed for refining the action region proposals. We validate the effectiveness of T3AL by conducting experiments on the THUMOS14 and the ActivityNet-v1.3 datasets. Our results demonstrate that T3AL significantly outperforms zero-shot baselines based on state-of-the-art VLMs, confirming the benefit of a test-time adaptation approach.
Temporal Action Detection (TAD) focuses on detecting pre-defined actions, while Moment Retrieval (MR) aims to identify the events described by open-ended natural language within untrimmed videos. Despite that they focus on different events, we observe they have a significant connection. For instance, most descriptions in MR involve multiple actions from TAD. In this paper, we aim to investigate the potential synergy between TAD and MR. Firstly, we propose a unified architecture, termed Unified Moment Detection (UniMD), for both TAD and MR. It transforms the inputs of the two tasks, namely actions for TAD or events for MR, into a common embedding space, and utilizes two novel query-dependent decoders to generate a uniform output of classification score and temporal segments. Secondly, we explore the efficacy of two task fusion learning approaches, pre-training and co-training, in order to enhance the mutual benefits between TAD and MR. Extensive experiments demonstrate that the proposed task fusion learning scheme enables the two tasks to help each other and outperform the separately trained counterparts. Impressively, UniMD achieves state-of-the-art results on three paired datasets Ego4D, Charades-STA, and ActivityNet. Our code is available at https://github.com/yingsen1/UniMD.
Video localization tasks aim to temporally locate specific instances in videos, including temporal action localization (TAL), sound event detection (SED) and audio-visual event localization (AVEL). Existing methods over-specialize on each task, overlooking the fact that these instances often occur in the same video to form the complete video content. In this work, we present UniAV, a Unified Audio-Visual perception network, to achieve joint learning of TAL, SED and AVEL tasks for the first time. UniAV can leverage diverse data available in task-specific datasets, allowing the model to learn and share mutually beneficial knowledge across tasks and modalities. To tackle the challenges posed by substantial variations in datasets (size/domain/duration) and distinct task characteristics, we propose to uniformly encode visual and audio modalities of all videos to derive generic representations, while also designing task-specific experts to capture unique knowledge for each task. Besides, we develop a unified language-aware classifier by utilizing a pre-trained text encoder, enabling the model to flexibly detect various types of instances and previously unseen ones by simply changing prompts during inference. UniAV outperforms its single-task counterparts by a large margin with fewer parameters, achieving on-par or superior performances compared to state-of-the-art task-specific methods across ActivityNet 1.3, DESED and UnAV-100 benchmarks.
While neural networks have excelled in video action recognition tasks, their black-box nature often obscures the understanding of their decision-making processes. Recent approaches used inherently interpretable models to analyze video actions in a manner akin to human reasoning. These models, however, usually fall short in performance compared to their black-box counterparts. In this work, we present a new framework named Language-guided Interpretable Action Recognition framework (LaIAR). LaIAR leverages knowledge from language models to enhance both the recognition capabilities and the interpretability of video models. In essence, we redefine the problem of understanding video model decisions as a task of aligning video and language models. Using the logical reasoning captured by the language model, we steer the training of the video model. This integrated approach not only improves the video model's adaptability to different domains but also boosts its overall performance. Extensive experiments on two complex video action datasets, Charades & CAD-120, validates the improved performance and interpretability of our LaIAR framework. The code of LaIAR is available at https://github.com/NingWang2049/LaIAR.
We introduce InternVideo2, a new family of video foundation models (ViFM) that achieve the state-of-the-art results in video recognition, video-text tasks, and video-centric dialogue. Our core design is a progressive training approach that unifies the masked video modeling, crossmodal contrastive learning, and next token prediction, scaling up the video encoder size to 6B parameters. At the data level, we prioritize spatiotemporal consistency by semantically segmenting videos and generating video-audio-speech captions. This improves the alignment between video and text. Through extensive experiments, we validate our designs and demonstrate superior performance on over 60 video and audio tasks. Notably, our model outperforms others on various video-related dialogue and long video understanding benchmarks, highlighting its ability to reason and comprehend longer contexts. Code and models are available at https://github.com/OpenGVLab/InternVideo/tree/main/InternVideo2/.
Manifold-valued measurements exist in numerous applications within computer vision and machine learning. Recent studies have extended Deep Neural Networks (DNNs) to manifolds, and concomitantly, normalization techniques have also been adapted to several manifolds, referred to as Riemannian normalization. Nonetheless, most of the existing Riemannian normalization methods have been derived in an ad hoc manner and only apply to specific manifolds. This paper establishes a unified framework for Riemannian Batch Normalization (RBN) techniques on Lie groups. Our framework offers the theoretical guarantee of controlling both the Riemannian mean and variance. Empirically, we focus on Symmetric Positive Definite (SPD) manifolds, which possess three distinct types of Lie group structures. Using the deformation concept, we generalize the existing Lie groups on SPD manifolds into three families of parameterized Lie groups. Specific normalization layers induced by these Lie groups are then proposed for SPD neural networks. We demonstrate the effectiveness of our approach through three sets of experiments: radar recognition, human action recognition, and electroencephalography (EEG) classification. The code is available at https://github.com/GitZH-Chen/LieBN.git.
Understanding human actions from body poses is critical for assistive robots sharing space with humans in order to make informed and safe decisions about the next interaction. However, precise temporal localization and annotation of activity sequences is time-consuming and the resulting labels are often noisy. If not effectively addressed, label noise negatively affects the model's training, resulting in lower recognition quality. Despite its importance, addressing label noise for skeleton-based action recognition has been overlooked so far. In this study, we bridge this gap by implementing a framework that augments well-established skeleton-based human action recognition methods with label-denoising strategies from various research areas to serve as the initial benchmark. Observations reveal that these baselines yield only marginal performance when dealing with sparse skeleton data. Consequently, we introduce a novel methodology, NoiseEraSAR, which integrates global sample selection, co-teaching, and Cross-Modal Mixture-of-Experts (CM-MOE) strategies, aimed at mitigating the adverse impacts of label noise. Our proposed approach demonstrates better performance on the established benchmark, setting new state-of-the-art standards. The source code for this study is accessible at https://github.com/xuyizdby/NoiseEraSAR.
Understanding videos is one of the fundamental directions in computer vision research, with extensive efforts dedicated to exploring various architectures such as RNN, 3D CNN, and Transformers. The newly proposed architecture of state space model, e.g., Mamba, shows promising traits to extend its success in long sequence modeling to video modeling. To assess whether Mamba can be a viable alternative to Transformers in the video understanding domain, in this work, we conduct a comprehensive set of studies, probing different roles Mamba can play in modeling videos, while investigating diverse tasks where Mamba could exhibit superiority. We categorize Mamba into four roles for modeling videos, deriving a Video Mamba Suite composed of 14 models/modules, and evaluating them on 12 video understanding tasks. Our extensive experiments reveal the strong potential of Mamba on both video-only and video-language tasks while showing promising efficiency-performance trade-offs. We hope this work could provide valuable data points and insights for future research on video understanding. Code is public: https://github.com/OpenGVLab/video-mamba-suite.
Adversarial examples generated by a surrogate model typically exhibit limited transferability to unknown target systems. To address this problem, many transferability enhancement approaches (e.g., input transformation and model augmentation) have been proposed. However, they show poor performances in attacking systems having different model genera from the surrogate model. In this paper, we propose a novel and generic attacking strategy, called Deformation-Constrained Warping Attack (DeCoWA), that can be effectively applied to cross model genus attack. Specifically, DeCoWA firstly augments input examples via an elastic deformation, namely Deformation-Constrained Warping (DeCoW), to obtain rich local details of the augmented input. To avoid severe distortion of global semantics led by random deformation, DeCoW further constrains the strength and direction of the warping transformation by a novel adaptive control strategy. Extensive experiments demonstrate that the transferable examples crafted by our DeCoWA on CNN surrogates can significantly hinder the performance of Transformers (and vice versa) on various tasks, including image classification, video action recognition, and audio recognition. Code is made available at https://github.com/LinQinLiang/DeCoWA.
Despite recent advances in video action recognition achieving strong performance on existing benchmarks, these models often lack robustness when faced with natural distribution shifts between training and test data. We propose two novel evaluation methods to assess model resilience to such distribution disparity. One method uses two different datasets collected from different sources and uses one for training and validation, and the other for testing. More precisely, we created dataset splits of HMDB-51 or UCF-101 for training, and Kinetics-400 for testing, using the subset of the classes that are overlapping in both train and test datasets. The other proposed method extracts the feature mean of each class from the target evaluation dataset's training data (i.e. class prototype) and estimates test video prediction as a cosine similarity score between each sample to the class prototypes of each target class. This procedure does not alter model weights using the target dataset and it does not require aligning overlapping classes of two different datasets, thus is a very efficient method to test the model robustness to distribution shifts without prior knowledge of the target distribution. We address the robustness problem by adversarial augmentation training - generating augmented views of videos that are "hard" for the classification model by applying gradient ascent on the augmentation parameters - as well as "curriculum" scheduling the strength of the video augmentations. We experimentally demonstrate the superior performance of the proposed adversarial augmentation approach over baselines across three state-of-the-art action recognition models - TSM, Video Swin Transformer, and Uniformer. The presented work provides critical insight into model robustness to distribution shifts and presents effective techniques to enhance video action recognition performance in a real-world deployment.
This paper proposes a framework that combines online human state estimation, action recognition and motion prediction to enable early assessment and prevention of worker biomechanical risk during lifting tasks. The framework leverages the NIOSH index to perform online risk assessment, thus fitting real-time applications. In particular, the human state is retrieved via inverse kinematics/dynamics algorithms from wearable sensor data. Human action recognition and motion prediction are achieved by implementing an LSTM-based Guided Mixture of Experts architecture, which is trained offline and inferred online. With the recognized actions, a single lifting activity is divided into a series of continuous movements and the Revised NIOSH Lifting Equation can be applied for risk assessment. Moreover, the predicted motions enable anticipation of future risks. A haptic actuator, embedded in the wearable system, can alert the subject of potential risk, acting as an active prevention device. The performance of the proposed framework is validated by executing real lifting tasks, while the subject is equipped with the iFeel wearable system.
Skeleton-based action recognition is a central task in human-computer interaction. However, most previous methods suffer from two issues: (i) semantic ambiguity arising from spatial-temporal information mixture; and (ii) overlooking the explicit exploitation of the latent data distributions (i.e., the intra-class variations and inter-class relations), thereby leading to sub-optimum solutions of the skeleton encoders. To mitigate this, we propose a spatial-temporal decoupling contrastive learning (STD-CL) framework to obtain discriminative and semantically distinct representations from the sequences, which can be incorporated into various previous skeleton encoders and can be removed when testing. Specifically, we decouple the global features into spatial-specific and temporal-specific features to reduce the spatial-temporal coupling of features. Furthermore, to explicitly exploit the latent data distributions, we employ the attentive features to contrastive learning, which models the cross-sequence semantic relations by pulling together the features from the positive pairs and pushing away the negative pairs. Extensive experiments show that STD-CL with four various skeleton encoders (HCN, 2S-AGCN, CTR-GCN, and Hyperformer) achieves solid improvements on NTU60, NTU120, and NW-UCLA benchmarks. The code will be released soon.
Temporal Action Localization (TAL) is a complex task that poses relevant challenges, particularly when attempting to generalize on new -- unseen -- domains in real-world applications. These scenarios, despite realistic, are often neglected in the literature, exposing these solutions to important performance degradation. In this work, we tackle this issue by introducing, for the first time, an approach for Unsupervised Domain Adaptation (UDA) in sparse TAL, which we refer to as Semantic Adversarial unsupervised Domain Adaptation (SADA). Our contributions are threefold: (1) we pioneer the development of a domain adaptation model that operates on realistic sparse action detection benchmarks; (2) we tackle the limitations of global-distribution alignment techniques by introducing a novel adversarial loss that is sensitive to local class distributions, ensuring finer-grained adaptation; and (3) we present a novel set of benchmarks based on EpicKitchens100 and CharadesEgo, that evaluate multiple domain shifts in a comprehensive manner. Our experiments indicate that SADA improves the adaptation across domains when compared to fully supervised state-of-the-art and alternative UDA methods, attaining a performance boost of up to 6.14% mAP.
Weakly-supervised temporal action localization aims to localize action instances in videos with only video-level action labels. Existing methods mainly embrace a localization-by-classification pipeline that optimizes the snippet-level prediction with a video classification loss. However, this formulation suffers from the discrepancy between classification and detection, resulting in inaccurate separation of foreground and background (F\&B) snippets. To alleviate this problem, we propose to explore the underlying structure among the snippets by resorting to unsupervised snippet clustering, rather than heavily relying on the video classification loss. Specifically, we propose a novel clustering-based F\&B separation algorithm. It comprises two core components: a snippet clustering component that groups the snippets into multiple latent clusters and a cluster classification component that further classifies the cluster as foreground or background. As there are no ground-truth labels to train these two components, we introduce a unified self-labeling mechanism based on optimal transport to produce high-quality pseudo-labels that match several plausible prior distributions. This ensures that the cluster assignments of the snippets can be accurately associated with their F\&B labels, thereby boosting the F\&B separation. We evaluate our method on three benchmarks: THUMOS14, ActivityNet v1.2 and v1.3. Our method achieves promising performance on all three benchmarks while being significantly more lightweight than previous methods. Code is available at https://github.com/Qinying-Liu/CASE
We propose a novel multimodal video benchmark - the Perception Test - to evaluate the perception and reasoning skills of pre-trained multimodal models (e.g. Flamingo, SeViLA, or GPT-4). Compared to existing benchmarks that focus on computational tasks (e.g. classification, detection or tracking), the Perception Test focuses on skills (Memory, Abstraction, Physics, Semantics) and types of reasoning (descriptive, explanatory, predictive, counterfactual) across video, audio, and text modalities, to provide a comprehensive and efficient evaluation tool. The benchmark probes pre-trained models for their transfer capabilities, in a zero-shot / few-shot or limited finetuning regime. For these purposes, the Perception Test introduces 11.6k real-world videos, 23s average length, designed to show perceptually interesting situations, filmed by around 100 participants worldwide. The videos are densely annotated with six types of labels (multiple-choice and grounded video question-answers, object and point tracks, temporal action and sound segments), enabling both language and non-language evaluations. The fine-tuning and validation splits of the benchmark are publicly available (CC-BY license), in addition to a challenge server with a held-out test split. Human baseline results compared to state-of-the-art video QA models show a substantial gap in performance (91.4% vs 46.2%), suggesting that there is significant room for improvement in multimodal video understanding. Dataset, baseline code, and challenge server are available at https://github.com/deepmind/perception_test
Knowledge distillation (KD), a technique widely employed in computer vision, has emerged as a de facto standard for improving the performance of small neural networks. However, prevailing KD-based approaches in video tasks primarily focus on designing loss functions and fusing cross-modal information. This overlooks the spatial-temporal feature semantics, resulting in limited advancements in model compression. Addressing this gap, our paper introduces an innovative knowledge distillation framework, with the generative model for training a lightweight student model. In particular, the framework is organized into two steps: the initial phase is Feature Representation, wherein a generative model-based attention module is trained to represent feature semantics; Subsequently, the Generative-based Feature Distillation phase encompasses both Generative Distillation and Attention Distillation, with the objective of transferring attention-based feature semantics with the generative model. The efficacy of our approach is demonstrated through comprehensive experiments on diverse popular datasets, proving considerable enhancements in video action recognition task. Moreover, the effectiveness of our proposed framework is validated in the context of more intricate video action detection task. Our code is available at https://github.com/aaai-24/Generative-based-KD.
Spatiotemporal graph convolutional networks (STGCNs) have emerged as a desirable model for skeleton-based human action recognition. Despite achieving state-of-the-art performance, there is a limited understanding of the representations learned by these models, which hinders their application in critical and real-world settings. While layerwise analysis of CNN models has been studied in the literature, to the best of our knowledge, there exists no study on the layerwise explainability of the embeddings learned on spatiotemporal data using STGCNs. In this paper, we first propose to use a local Dataset Graph (DS-Graph) obtained from the feature representation of input data at each layer to develop an understanding of the layer-wise embedding geometry of the STGCN. To do so, we develop a window-based dynamic time warping (DTW) method to compute the distance between data sequences with varying temporal lengths. To validate our findings, we have developed a layer-specific Spatiotemporal Graph Gradient-weighted Class Activation Mapping (L-STG-GradCAM) technique tailored for spatiotemporal data. This approach enables us to visually analyze and interpret each layer within the STGCN network. We characterize the functions learned by each layer of the STGCN using the label smoothness of the representation and visualize them using our L-STG-GradCAM approach. Our proposed method is generic and can yield valuable insights for STGCN architectures in different applications. However, this paper focuses on the human activity recognition task as a representative application. Our experiments show that STGCN models learn representations that capture general human motion in their initial layers while discriminating different actions only in later layers. This justifies experimental observations showing that fine-tuning deeper layers works well for transfer between related tasks.