This paper presents a study on improving human action recognition through the utilization of knowledge distillation, and the combination of CNN and ViT models. The research aims to enhance the performance and efficiency of smaller student models by transferring knowledge from larger teacher models. The proposed method employs a Transformer vision network as the student model, while a convolutional network serves as the teacher model. The teacher model extracts local image features, whereas the student model focuses on global features using an attention mechanism. The Vision Transformer (ViT) architecture is introduced as a robust framework for capturing global dependencies in images. Additionally, advanced variants of ViT, namely PVT, Convit, MVIT, Swin Transformer, and Twins, are discussed, highlighting their contributions to computer vision tasks. The ConvNeXt model is introduced as a teacher model, known for its efficiency and effectiveness in computer vision. The paper presents performance results for human action recognition on the Stanford 40 dataset, comparing the accuracy and mAP of student models trained with and without knowledge distillation. The findings illustrate that the suggested approach significantly improves the accuracy and mAP when compared to training networks under regular settings. These findings emphasize the potential of combining local and global features in action recognition tasks.
Human action recognition from skeletal data is an important and active area of research in which the state of the art has not yet achieved near-perfect accuracy on many well-known datasets. In this paper, we introduce the Distribution of Action Movements Descriptor, a novel action descriptor based on the distribution of the directions of the motions of the joints between frames, over the set of all possible motions in the dataset. The descriptor is computed as a normalized histogram over a set of representative directions of the joints, which are in turn obtained via clustering. While the descriptor is global in the sense that it represents the overall distribution of movement directions of an action, it is able to partially retain its temporal structure by applying a windowing scheme. The descriptor, together with a standard classifier, outperforms several state-of-the-art techniques on many well-known datasets.
Vision-language models (VLMs) classify the query video by calculating a similarity score between the visual features and text-based class label representations. Recently, large language models (LLMs) have been used to enrich the text-based class labels by enhancing the descriptiveness of the class names. However, these improvements are restricted to the text-based classifier only, and the query visual features are not considered. In this paper, we propose a framework which combines pre-trained discriminative VLMs with pre-trained generative video-to-text and text-to-text models. We introduce two key modifications to the standard zero-shot setting. First, we propose language-guided visual feature enhancement and employ a video-to-text model to convert the query video to its descriptive form. The resulting descriptions contain vital visual cues of the query video, such as what objects are present and their spatio-temporal interactions. These descriptive cues provide additional semantic knowledge to VLMs to enhance their zeroshot performance. Second, we propose video-specific prompts to LLMs to generate more meaningful descriptions to enrich class label representations. Specifically, we introduce prompt techniques to create a Tree Hierarchy of Categories for class names, offering a higher-level action context for additional visual cues, We demonstrate the effectiveness of our approach in video understanding across three different zero-shot settings: 1) video action recognition, 2) video-to-text and textto-video retrieval, and 3) time-sensitive video tasks. Consistent improvements across multiple benchmarks and with various VLMs demonstrate the effectiveness of our proposed framework. Our code will be made publicly available.
Human Action Recognition plays a driving engine of many human-computer interaction applications. Most current researches focus on improving the model generalization by integrating multiple homogeneous modalities, including RGB images, human poses, and optical flows. Furthermore, contextual interactions and out-of-context sign languages have been validated to depend on scene category and human per se. Those attempts to integrate appearance features and human poses have shown positive results. However, with human poses’ spatial errors and temporal ambiguities, existing methods are subject to poor scalability, limited robustness, and sub-optimal models. In this paper, inspired by the assumption that different modalities may maintain temporal consistency and spatial complementarity, we present a novel Bi-directional Co-temporal and Cross-spatial Attention Fusion Model (B2C-AFM). Our model is characterized by the asynchronous fusion strategy of multi-modal features along temporal and spatial dimensions. Besides, the novel explicit motion-oriented pose representations called Limb Flow Fields (Lff) are explored to alleviate the temporal ambiguity regarding human poses. Experiments on publicly available datasets validate our contributions. Abundant ablation studies experimentally show that B2C-AFM achieves robust performance across seen and unseen human actions. The codes are available at https://github.com/gftww/B2C.git.
This paper tackles the challenge of point-supervised temporal action detection, wherein only a single frame is annotated for each action instance in the training set. Most of the current methods, hindered by the sparse nature of annotated points, struggle to effectively represent the continuous structure of actions or the inherent temporal and semantic dependencies within action instances. Consequently, these methods frequently learn merely the most distinctive segments of actions, leading to the creation of incomplete action proposals. This paper proposes POTLoc, a Pseudo-label Oriented Transformer for weakly-supervised Action Localization utilizing only point-level annotation. POTLoc is designed to identify and track continuous action structures via a self-training strategy. The base model begins by generating action proposals solely with point-level supervision. These proposals undergo refinement and regression to enhance the precision of the estimated action boundaries, which subsequently results in the production of `pseudo-labels' to serve as supplementary supervisory signals. The architecture of the model integrates a transformer with a temporal feature pyramid to capture video snippet dependencies and model actions of varying duration. The pseudo-labels, providing information about the coarse locations and boundaries of actions, assist in guiding the transformer for enhanced learning of action dynamics. POTLoc outperforms the state-of-the-art point-supervised methods on THUMOS'14 and ActivityNet-v1.2 datasets.
Video instance segmentation, also known as multi-object tracking and segmentation, is an emerging computer vision research area introduced in 2019, aiming at detecting, segmenting, and tracking instances in videos simultaneously. By tackling the video instance segmentation tasks through effective analysis and utilization of visual information in videos, a range of computer vision-enabled applications (e.g., human action recognition, medical image processing, autonomous vehicle navigation, surveillance, etc) can be implemented. As deep-learning techniques take a dominant role in various computer vision areas, a plethora of deep-learning-based video instance segmentation schemes have been proposed. This survey offers a multifaceted view of deep-learning schemes for video instance segmentation, covering various architectural paradigms, along with comparisons of functional performance, model complexity, and computational overheads. In addition to the common architectural designs, auxiliary techniques for improving the performance of deep-learning models for video instance segmentation are compiled and discussed. Finally, we discuss a range of major challenges and directions for further investigations to help advance this promising research field.
Point-level supervised temporal action localization (PTAL) aims at recognizing and localizing actions in untrimmed videos where only a single point (frame) within every action instance is annotated in training data. Without temporal annotations, most previous works adopt the multiple instance learning (MIL) framework, where the input video is segmented into non-overlapped short snippets, and action classification is performed independently on every short snippet. We argue that the MIL framework is suboptimal for PTAL because it operates on separated short snippets that contain limited temporal information. Therefore, the classifier only focuses on several easy-to-distinguish snippets instead of discovering the whole action instance without missing any relevant snippets. To alleviate this problem, we propose a novel method that localizes actions by generating and evaluating action proposals of flexible duration that involve more comprehensive temporal information. Moreover, we introduce an efficient clustering algorithm to efficiently generate dense pseudo labels that provide stronger supervision, and a fine-grained contrastive loss to further refine the quality of pseudo labels. Experiments show that our proposed method achieves competitive or superior performance to the state-of-the-art methods and some fully-supervised methods on four benchmarks: ActivityNet 1.3, THUMOS 14, GTEA, and BEOID datasets.
Foundational multimodal models pre-trained on large scale image-text pairs or video-text pairs or both have shown strong generalization abilities on downstream tasks. However unlike image-text models, pretraining video-text models is always not feasible due to the difficulty in collecting large-scale clean and aligned data, and exponential computational costs involved in the pretraining phase. Therefore, the pertinent question to ask is: Can image-text models be adapted to video tasks and is there any benefit to using these models over pretraining directly on videos? In this work, we focus on this question by proposing a detailed study on the generalization abilities of image-text models when evaluated on video understanding tasks in a zero-shot setting. We investigate 9 foundational image-text models on a diverse set of video tasks that include video action recognition (video AR), video retrieval (video RT), video question answering (video QA), video multiple choice (video MC) and video captioning (video CP). Our experiments show that image-text models exhibit impressive performance on video AR, video RT and video MC. Furthermore, they perform moderately on video captioning and poorly on video QA. These findings shed a light on the benefits of adapting foundational image-text models to an array of video tasks while avoiding the costly pretraining step.
Temporal Action Localization (TAL) aims to identify actions' start, end, and class labels in untrimmed videos. While recent advancements using transformer networks and Feature Pyramid Networks (FPN) have enhanced visual feature recognition in TAL tasks, less progress has been made in the integration of audio features into such frameworks. This paper introduces the Multi-Resolution Audio-Visual Feature Fusion (MRAV-FF), an innovative method to merge audio-visual data across different temporal resolutions. Central to our approach is a hierarchical gated cross-attention mechanism, which discerningly weighs the importance of audio information at diverse temporal scales. Such a technique not only refines the precision of regression boundaries but also bolsters classification confidence. Importantly, MRAV-FF is versatile, making it compatible with existing FPN TAL architectures and offering a significant enhancement in performance when audio data is available.
We present a new pre-training strategy called M$^{3}$3D ($\underline{M}$ulti-$\underline{M}$odal $\underline{M}$asked $\underline{3D}$) built based on Multi-modal masked autoencoders that can leverage 3D priors and learned cross-modal representations in RGB-D data. We integrate two major self-supervised learning frameworks; Masked Image Modeling (MIM) and contrastive learning; aiming to effectively embed masked 3D priors and modality complementary features to enhance the correspondence between modalities. In contrast to recent approaches which are either focusing on specific downstream tasks or require multi-view correspondence, we show that our pre-training strategy is ubiquitous, enabling improved representation learning that can transfer into improved performance on various downstream tasks such as video action recognition, video action detection, 2D semantic segmentation and depth estimation. Experiments show that M$^{3}$3D outperforms the existing state-of-the-art approaches on ScanNet, NYUv2, UCF-101 and OR-AR, particularly with an improvement of +1.3\% mIoU against Mask3D on ScanNet semantic segmentation. We further evaluate our method on low-data regime and demonstrate its superior data efficiency compared to current state-of-the-art approaches.
The goal of Temporal Action Localization (TAL) is to find the categories and temporal boundaries of actions in an untrimmed video. Most TAL methods rely heavily on action recognition models that are sensitive to action labels rather than temporal boundaries. More importantly, few works consider the background frames that are similar to action frames in pixels but dissimilar in semantics, which also leads to inaccurate temporal boundaries. To address the challenge above, we propose a Boundary-Aware Proposal Generation (BAPG) method with contrastive learning. Specifically, we define the above background frames as hard negative samples. Contrastive learning with hard negative mining is introduced to improve the discrimination of BAPG. BAPG is independent of the existing TAL network architecture, so it can be applied plug-and-play to mainstream TAL models. Extensive experimental results on THUMOS14 and ActivityNet-1.3 demonstrate that BAPG can significantly improve the performance of TAL.
To integrate action recognition into autonomous robotic systems, it is essential to address challenges such as person occlusions-a common yet often overlooked scenario in existing self-supervised skeleton-based action recognition methods. In this work, we propose IosPSTL, a simple and effective self-supervised learning framework designed to handle occlusions. IosPSTL combines a cluster-agnostic KNN imputer with an Occluded Partial Spatio-Temporal Learning (OPSTL) strategy. First, we pre-train the model on occluded skeleton sequences. Then, we introduce a cluster-agnostic KNN imputer that performs semantic grouping using k-means clustering on sequence embeddings. It imputes missing skeleton data by applying K-Nearest Neighbors in the latent space, leveraging nearby sample representations to restore occluded joints. This imputation generates more complete skeleton sequences, which significantly benefits downstream self-supervised models. To further enhance learning, the OPSTL module incorporates Adaptive Spatial Masking (ASM) to make better use of intact, high-quality skeleton sequences during training. Our method achieves state-of-the-art performance on the occluded versions of the NTU-60 and NTU-120 datasets, demonstrating its robustness and effectiveness under challenging conditions. Code is available at https://github.com/cyfml/OPSTL.
Self-supervised representation learning for human action recognition has developed rapidly in recent years. Most of the existing works are based on skeleton data while using a multi-modality setup. These works overlooked the differences in performance among modalities, which led to the propagation of erroneous knowledge between modalities while only three fundamental modalities, i.e., joints, bones, and motions are used, hence no additional modalities are explored. In this work, we first propose an Implicit Knowledge Exchange Module (IKEM) which alleviates the propagation of erroneous knowledge between low-performance modalities. Then, we further propose three new modalities to enrich the complementary information between modalities. Finally, to maintain efficiency when introducing new modalities, we propose a novel teacher-student framework to distill the knowledge from the secondary modalities into the mandatory modalities considering the relationship constrained by anchors, positives, and negatives, named relational cross-modality knowledge distillation. The experimental results demonstrate the effectiveness of our approach, unlocking the efficient use of skeleton-based multi-modality data. Source code will be made publicly available at https://github.com/desehuileng0o0/IKEM.
While most modern video understanding models operate on short-range clips, real-world videos are often several minutes long with semantically consistent segments of variable length. A common approach to process long videos is applying a short-form video model over uniformly sampled clips of fixed temporal length and aggregating the outputs. This approach neglects the underlying nature of long videos since fixed-length clips are often redundant or uninformative. In this paper, we aim to provide a generic and adaptive sampling approach for long-form videos in lieu of the de facto uniform sampling. Viewing videos as semantically consistent segments, we formulate a task-agnostic, unsupervised, and scalable approach based on Kernel Temporal Segmentation (KTS) for sampling and tokenizing long videos. We evaluate our method on long-form video understanding tasks such as video classification and temporal action localization, showing consistent gains over existing approaches and achieving state-of-the-art performance on long-form video modeling.
Action scene understanding in soccer is a challenging task due to the complex and dynamic nature of the game, as well as the interactions between players. This article provides a comprehensive overview of this task divided into action recognition, spotting, and spatio-temporal action localization, with a particular emphasis on the modalities used and multimodal methods. We explore the publicly available data sources and metrics used to evaluate models' performance. The article reviews recent state-of-the-art methods that leverage deep learning techniques and traditional methods. We focus on multimodal methods, which integrate information from multiple sources, such as video and audio data, and also those that represent one source in various ways. The advantages and limitations of methods are discussed, along with their potential for improving the accuracy and robustness of models. Finally, the article highlights some of the open research questions and future directions in the field of soccer action recognition, including the potential for multimodal methods to advance this field. Overall, this survey provides a valuable resource for researchers interested in the field of action scene understanding in soccer.
The recent advances in Convolutional Neural Networks (CNNs) and Vision Transformers have convincingly demonstrated high learning capability for video action recognition on large datasets. Nevertheless, deep models often suffer from the overfitting effect on small-scale datasets with a limited number of training videos. A common solution is to exploit the existing image augmentation strategies for each frame individually including Mixup, Cutmix, and RandAugment, which are not particularly optimized for video data. In this paper, we propose a novel video augmentation strategy named Selective Volume Mixup (SV-Mix) to improve the generalization ability of deep models with limited training videos. SV-Mix devises a learnable selective module to choose the most informative volumes from two videos and mixes the volumes up to achieve a new training video. Technically, we propose two new modules, i.e., a spatial selective module to select the local patches for each spatial position, and a temporal selective module to mix the entire frames for each timestamp and maintain the spatial pattern. At each time, we randomly choose one of the two modules to expand the diversity of training samples. The selective modules are jointly optimized with the video action recognition framework to find the optimal augmentation strategy. We empirically demonstrate the merits of the SV-Mix augmentation on a wide range of video action recognition benchmarks and consistently boot the performances of both CNN-based and transformer-based models.
Point-level weakly-supervised temporal action localization (PWTAL) aims to localize actions with only a single timestamp annotation for each action instance. Existing methods tend to mine dense pseudo labels to alleviate the label sparsity, but overlook the potential sub-action temporal structures, resulting in inferior performance. To tackle this problem, we propose a novel sub-action prototype learning framework (SPL-Loc) which comprises Sub-action Prototype Clustering (SPC) and Ordered Prototype Alignment (OPA). SPC adaptively extracts representative sub-action prototypes which are capable to perceive the temporal scale and spatial content variation of action instances. OPA selects relevant prototypes to provide completeness clue for pseudo label generation by applying a temporal alignment loss. As a result, pseudo labels are derived from alignment results to improve action boundary prediction. Extensive experiments on three popular benchmarks demonstrate that the proposed SPL-Loc significantly outperforms existing SOTA PWTAL methods.
Action recognition is a key technology for many industrial applications. Methods using visual information such as images are very popular. However, privacy issues prevent widespread usage due to the inclusion of private information, such as visible faces and scene backgrounds, which are not necessary for recognizing user action. In this paper, we propose a privacy-preserving action recognition by ultrasound active sensing. As action recognition from ultrasound active sensing in a non-invasive manner is not well investigated, we create a new dataset for action recognition and conduct a comparison of features for classification. We calculated feature values by focusing on the temporal variation of the amplitude of ultrasound reflected waves and performed classification using a support vector machine and VGG for eight fundamental action classes. We confirmed that our method achieved an accuracy of 97.9% when trained and evaluated on the same person and in the same environment. Additionally, our method achieved an accuracy of 89.5% even when trained and evaluated on different people. We also report the analyses of accuracies in various conditions and limitations.
Human action recognition (HAR) is a high-level and significant research area in computer vision due to its ubiquitous applications. The main limitations of the current HAR models are their complex structures and lengthy training time. In this paper, we propose a simple yet versatile and effective end-to-end deep learning architecture, coined as TransNet, for HAR. TransNet decomposes the complex 3D-CNNs into 2D- and 1D-CNNs, where the 2D- and 1D-CNN components extract spatial features and temporal patterns in videos, respectively. Benefiting from its concise architecture, TransNet is ideally compatible with any pretrained state-of-the-art 2D-CNN models in other fields, being transferred to serve the HAR task. In other words, it naturally leverages the power and success of transfer learning for HAR, bringing huge advantages in terms of efficiency and effectiveness. Extensive experimental results and the comparison with the state-of-the-art models demonstrate the superior performance of the proposed TransNet in HAR in terms of flexibility, model complexity, training speed and classification accuracy.
Various types of sensors have been considered to develop human action recognition (HAR) models. Robust HAR performance can be achieved by fusing multimodal data acquired by different sensors. In this paper, we introduce a new multimodal fusion architecture, referred to as Unified Contrastive Fusion Transformer (UCFFormer) designed to integrate data with diverse distributions to enhance HAR performance. Based on the embedding features extracted from each modality, UCFFormer employs the Unified Transformer to capture the inter-dependency among embeddings in both time and modality domains. We present the Factorized Time-Modality Attention to perform self-attention efficiently for the Unified Transformer. UCFFormer also incorporates contrastive learning to reduce the discrepancy in feature distributions across various modalities, thus generating semantically aligned features for information fusion. Performance evaluation conducted on two popular datasets, UTD-MHAD and NTU RGB+D, demonstrates that UCFFormer achieves state-of-the-art performance, outperforming competing methods by considerable margins.