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.
Recent advancements in large-scale pre-training of visual-language models on paired image-text data have demonstrated impressive generalization capabilities for zero-shot tasks. Building on this success, efforts have been made to adapt these image-based visual-language models, such as CLIP, for videos extending their zero-shot capabilities to the video domain. While these adaptations have shown promising results, they come at a significant computational cost and struggle with effectively modeling the crucial temporal aspects inherent to the video domain. In this study, we present EZ-CLIP, a simple and efficient adaptation of CLIP that addresses these challenges. EZ-CLIP leverages temporal visual prompting for seamless temporal adaptation, requiring no fundamental alterations to the core CLIP architecture while preserving its remarkable generalization abilities. Moreover, we introduce a novel learning objective that guides the temporal visual prompts to focus on capturing motion, thereby enhancing its learning capabilities from video data. We conducted extensive experiments on five different benchmark datasets, thoroughly evaluating EZ-CLIP for zero-shot learning and base-to-novel video action recognition, and also demonstrating its potential for few-shot generalization.Impressively, with a mere 5.2 million learnable parameters (as opposed to the 71.1 million in the prior best model), EZ-CLIP can be efficiently trained on a single GPU, outperforming existing approaches in several evaluations.
Recently, temporal action localization (TAL) has garnered significant interest in information retrieval community. However, existing supervised/weakly supervised methods are heavily dependent on extensive labeled temporal boundaries and action categories, which is labor-intensive and time-consuming. Although some unsupervised methods have utilized the ``iteratively clustering and localization'' paradigm for TAL, they still suffer from two pivotal impediments: 1) unsatisfactory video clustering confidence, and 2) unreliable video pseudolabels for model training. To address these limitations, we present a novel self-paced iterative learning model to enhance clustering and localization training simultaneously, thereby facilitating more effective unsupervised TAL. Concretely, we improve the clustering confidence through exploring the contextual feature-robust visual information. Thereafter, we design two (constant- and variable- speed) incremental instance learning strategies for easy-to-hard model training, thus ensuring the reliability of these video pseudolabels and further improving overall localization performance. Extensive experiments on two public datasets have substantiated the superiority of our model over several state-of-the-art competitors.
Mobile service robots are proving to be increasingly effective in a range of applications, such as healthcare, monitoring Activities of Daily Living (ADL), and facilitating Ambient Assisted Living (AAL). These robots heavily rely on Human Action Recognition (HAR) to interpret human actions and intentions. However, for HAR to function effectively on service robots, it requires prior knowledge of human presence (human detection) and identification of individuals to monitor (human tracking). In this work, we propose an end-to-end pipeline that encompasses the entire process, starting from human detection and tracking, leading to action recognition. The pipeline is designed to operate in near real-time while ensuring all stages of processing are performed on the edge, reducing the need for centralised computation. To identify the most suitable models for our mobile robot, we conducted a series of experiments comparing state-of-the-art solutions based on both their detection performance and efficiency. To evaluate the effectiveness of our proposed pipeline, we proposed a dataset comprising daily household activities. By presenting our findings and analysing the results, we demonstrate the efficacy of our approach in enabling mobile robots to understand and respond to human behaviour in real-world scenarios relying mainly on the data from their RGB cameras.
Exploring open-vocabulary video action recognition is a promising venture, which aims to recognize previously unseen actions within any arbitrary set of categories. Existing methods typically adapt pretrained image-text models to the video domain, capitalizing on their inherent strengths in generalization. A common thread among such methods is the augmentation of visual embeddings with temporal information to improve the recognition of seen actions. Yet, they compromise with standard less-informative action descriptions, thus faltering when confronted with novel actions. Drawing inspiration from human cognitive processes, we argue that augmenting text embeddings with human prior knowledge is pivotal for open-vocabulary video action recognition. To realize this, we innovatively blend video models with Large Language Models (LLMs) to devise Action-conditioned Prompts. Specifically, we harness the knowledge in LLMs to produce a set of descriptive sentences that contain distinctive features for identifying given actions. Building upon this foundation, we further introduce a multi-modal action knowledge alignment mechanism to align concepts in video and textual knowledge encapsulated within the prompts. Extensive experiments on various video benchmarks, including zero-shot, few-shot, and base-to-novel generalization settings, demonstrate that our method not only sets new SOTA performance but also possesses excellent interpretability.
In this work, we tackle the problem of unsupervised domain adaptation (UDA) for video action recognition. Our approach, which we call UNITE, uses an image teacher model to adapt a video student model to the target domain. UNITE first employs self-supervised pre-training to promote discriminative feature learning on target domain videos using a teacher-guided masked distillation objective. We then perform self-training on masked target data, using the video student model and image teacher model together to generate improved pseudolabels for unlabeled target videos. Our self-training process successfully leverages the strengths of both models to achieve strong transfer performance across domains. We evaluate our approach on multiple video domain adaptation benchmarks and observe significant improvements upon previously reported results.
Video recognition models often learn scene-biased action representation due to the spurious correlation between actions and scenes in the training data. Such models show poor performance when the test data consists of videos with unseen action-scene combinations. Although scene-debiased action recognition models might address the issue, they often overlook valuable scene information in the data. To address this challenge, we propose to learn DisEntangled VIdeo representations of Action and Scene (DEVIAS), for more holistic video understanding. We propose an encoder-decoder architecture to learn disentangled action and scene representations with a single model. The architecture consists of a disentangling encoder (DE), an action mask decoder (AMD), and a prediction head. The key to achieving the disentanglement is employing both DE and AMD during training time. The DE uses the slot attention mechanism to learn disentangled action and scene representations. For further disentanglement, an AMD learns to predict action masks, given an action slot. With the resulting disentangled representations, we can achieve robust performance across diverse scenarios, including both seen and unseen action-scene combinations. We rigorously validate the proposed method on the UCF-101, Kinetics-400, and HVU datasets for the seen, and the SCUBA, HAT, and HVU datasets for unseen action-scene combination scenarios. Furthermore, DEVIAS provides flexibility to adjust the emphasis on action or scene information depending on dataset characteristics for downstream tasks. DEVIAS shows favorable performance in various downstream tasks: Diving48, Something-Something-V2, UCF-101, and ActivityNet. The code is available at https://github.com/KHU-VLL/DEVIAS.
Pre-training on massive video datasets has become essential to achieve high action recognition performance on smaller downstream datasets. However, most large-scale video datasets contain images of people and hence are accompanied with issues related to privacy, ethics, and data protection, often preventing them from being publicly shared for reproducible research. Existing work has attempted to alleviate these problems by blurring faces, downsampling videos, or training on synthetic data. On the other hand, analysis on the transferability of privacy-preserving pre-trained models to downstream tasks has been limited. In this work, we study this problem by first asking the question: can we pre-train models for human action recognition with data that does not include real humans? To this end, we present, for the first time, a benchmark that leverages real-world videos with humans removed and synthetic data containing virtual humans to pre-train a model. We then evaluate the transferability of the representation learned on this data to a diverse set of downstream action recognition benchmarks. Furthermore, we propose a novel pre-training strategy, called Privacy-Preserving MAE-Align, to effectively combine synthetic data and human-removed real data. Our approach outperforms previous baselines by up to 5% and closes the performance gap between human and no-human action recognition representations on downstream tasks, for both linear probing and fine-tuning. Our benchmark, code, and models are available at https://github.com/howardzh01/PPMA .
Recently, temporal action detection (TAD) has seen significant performance improvement with end-to-end training. However, due to the memory bottleneck, only models with limited scales and limited data volumes can afford end-to-end training, which inevitably restricts TAD performance. In this paper, we reduce the memory consumption for end-to-end training, and manage to scale up the TAD backbone to 1 billion parameters and the input video to 1,536 frames, leading to significant detection performance. The key to our approach lies in our proposed temporal-informative adapter (TIA), which is a novel lightweight module that reduces training memory. Using TIA, we free the humongous backbone from learning to adapt to the TAD task by only updating the parameters in TIA. TIA also leads to better TAD representation by temporally aggregating context from adjacent frames throughout the backbone. We evaluate our model across four representative datasets. Owing to our efficient design, we are able to train end-to-end on VideoMAEv2-giant and achieve 75.4% mAP on THUMOS14, being the first end-to-end model to outperform the best feature-based methods. Code is available at https://github.com/sming256/AdaTAD.
Developing end-to-end action recognition models on long videos is fundamental and crucial for long-video action understanding. Due to the unaffordable cost of end-to-end training on the whole long videos, existing works generally train models on short clips trimmed from long videos. However, this ``trimming-then-training'' practice requires action interval annotations for clip-level supervision, i.e., knowing which actions are trimmed into the clips. Unfortunately, collecting such annotations is very expensive and prevents model training at scale. To this end, this work aims to build a weakly supervised end-to-end framework for training recognition models on long videos, with only video-level action category labels. Without knowing the precise temporal locations of actions in long videos, our proposed weakly supervised framework, namely AdaptFocus, estimates where and how likely the actions will occur to adaptively focus on informative action clips for end-to-end training. The effectiveness of the proposed AdaptFocus framework is demonstrated on three long-video datasets. Furthermore, for downstream long-video tasks, our AdaptFocus framework provides a weakly supervised feature extraction pipeline for extracting more robust long-video features, such that the state-of-the-art methods on downstream tasks are significantly advanced. We will release the code and models.
Video Moment Retrieval (MR) and Highlight Detection (HD) have attracted significant attention due to the growing demand for video analysis. Recent approaches treat MR and HD as similar video grounding problems and address them together with transformer-based architecture. However, we observe that the emphasis of MR and HD differs, with one necessitating the perception of local relationships and the other prioritizing the understanding of global contexts. Consequently, the lack of task-specific design will inevitably lead to limitations in associating the intrinsic specialty of two tasks. To tackle the issue, we propose a Unified Video COMprehension framework (UVCOM) to bridge the gap and jointly solve MR and HD effectively. By performing progressive integration on intra and inter-modality across multi-granularity, UVCOM achieves the comprehensive understanding in processing a video. Moreover, we present multi-aspect contrastive learning to consolidate the local relation modeling and global knowledge accumulation via well aligned multi-modal space. Extensive experiments on QVHighlights, Charades-STA, TACoS , YouTube Highlights and TVSum datasets demonstrate the effectiveness and rationality of UVCOM which outperforms the state-of-the-art methods by a remarkable margin.
Nowadays, humans and robots are working more closely together. This increases business productivity and product quality, leading to efficiency and growth. However, human and robot collaboration is rather static; robots move to a specific position then humans perform their tasks while being assisted by the robots. In order to get a dynamic collaboration, robots need to understand the human's intention and learn to recognize the performed actions complementing therefore his capabilities and relieving him of arduous tasks. Consequently, there is a need for a human action recognition dataset for Machine Learning algorithms. Currently available depth-based and RGB+D+S based human action recognition datasets have a number of limitations, counting the lack of training samples along with distinct class labels, camera views, diversity of subjects and more importantly the absence of actual industrial human actions in an industrial environment. Actual action recognition datasets include simple daily, mutual, or health-related actions. Therefore, in this paper we introduce an RGB+ S dataset named “Industrial Human Action Recognition Dataset” (InHARD) from a real-world setting for industrial human action recognition with over 2 million frames, collected from 16 distinct subjects. This dataset contains 13 different industrial action classes and over 4800 action samples. The introduction of this dataset should allow us the study and development of various learning techniques for the task of human actions analysis inside industrial environments involving human robot collaborations.
As of today, state-of-the-art activity recognition from wearable sensors relies on algorithms being trained to classify fixed windows of data. In contrast, video-based Human Activity Recognition, known as Temporal Action Localization (TAL), has followed a segment-based prediction approach, localizing activity segments in a timeline of arbitrary length. This paper is the first to systematically demonstrate the applicability of state-of-the-art TAL models for both offline and near-online Human Activity Recognition (HAR) using raw inertial data as well as pre-extracted latent features as input. Offline prediction results show that TAL models are able to outperform popular inertial models on a multitude of HAR benchmark datasets, with improvements reaching as much as 26% in F1-score. We show that by analyzing timelines as a whole, TAL models can produce more coherent segments and achieve higher NULL-class accuracy across all datasets. We demonstrate that TAL is less suited for the immediate classification of small-sized windows of data, yet offers an interesting perspective on inertial-based HAR -- alleviating the need for fixed-size windows and enabling algorithms to recognize activities of arbitrary length. With design choices and training concepts yet to be explored, we argue that TAL architectures could be of significant value to the inertial-based HAR community. The code and data download to reproduce experiments is publicly available via github.com/mariusbock/tal_for_har.
This paper addresses the challenge of point-supervised temporal action detection, in which only one frame per action instance is annotated in the training set. Self-training aims to provide supplementary supervision for the training process by generating pseudo-labels (action proposals) from a base model. However, most current methods generate action proposals by applying manually designed thresholds to action classification probabilities and treating adjacent snippets as independent entities. As a result, these methods struggle to generate complete action proposals, exhibit sensitivity to fluctuations in action classification scores, and generate redundant and overlapping action proposals. This paper proposes a novel framework termed ADM-Loc, which stands for Actionness Distribution Modeling for point-supervised action Localization. ADM-Loc generates action proposals by fitting a composite distribution, comprising both Gaussian and uniform distributions, to the action classification signals. This fitting process is tailored to each action class present in the video and is applied separately for each action instance, ensuring the distinctiveness of their distributions. ADM-Loc significantly enhances the alignment between the generated action proposals and ground-truth action instances and offers high-quality pseudo-labels for self-training. Moreover, to model action boundary snippets, it enforces consistency in action classification scores during training by employing Gaussian kernels, supervised with the proposed loss functions. ADM-Loc outperforms the state-of-the-art point-supervised methods on THUMOS14 and ActivityNet-v1.2 datasets.
Large-scale visual-language pre-trained models have achieved significant success in various video tasks. However, most existing methods follow an "adapt then align" paradigm, which adapts pre-trained image encoders to model video-level representations and utilizes one-hot or text embedding of the action labels for supervision. This paradigm overlooks the challenge of mapping from static images to complicated activity concepts. In this paper, we propose a novel "Align before Adapt" (ALT) paradigm. Prior to adapting to video representation learning, we exploit the entity-to-region alignments for each frame. The alignments are fulfilled by matching the region-aware image embeddings to an offline-constructed text corpus. With the aligned entities, we feed their text embeddings to a transformer-based video adapter as the queries, which can help extract the semantics of the most important entities from a video to a vector. This paradigm reuses the visual-language alignment of VLP during adaptation and tries to explain an action by the underlying entities. This helps understand actions by bridging the gap with complex activity semantics, particularly when facing unfamiliar or unseen categories. ALT demonstrates competitive performance while maintaining remarkably low computational costs. In fully supervised experiments, it achieves 88.1% top-1 accuracy on Kinetics-400 with only 4947 GFLOPs. Moreover, ALT outperforms the previous state-of-the-art methods in both zero-shot and few-shot experiments, emphasizing its superior generalizability across various learning scenarios.
Automated human action recognition, a burgeoning field within computer vision, boasts diverse applications spanning surveillance, security, human-computer interaction, tele-health, and sports analysis. Precise action recognition in infants serves a multitude of pivotal purposes, encompassing safety monitoring, developmental milestone tracking, early intervention for developmental delays, fostering parent-infant bonds, advancing computer-aided diagnostics, and contributing to the scientific comprehension of child development. This paper delves into the intricacies of infant action recognition, a domain that has remained relatively uncharted despite the accomplishments in adult action recognition. In this study, we introduce a groundbreaking dataset called ``InfActPrimitive'', encompassing five significant infant milestone action categories, and we incorporate specialized preprocessing for infant data. We conducted an extensive comparative analysis employing cutting-edge skeleton-based action recognition models using this dataset. Our findings reveal that, although the PoseC3D model achieves the highest accuracy at approximately 71%, the remaining models struggle to accurately capture the dynamics of infant actions. This highlights a substantial knowledge gap between infant and adult action recognition domains and the urgent need for data-efficient pipeline models.
Impressed by the coolest skateboarding sports program from 2021 Tokyo Olympic Games, we are the first to curate the original real-world video datasets "SkateboardAI" in the wild, even self-design and implement diverse uni-modal and multi-modal video action recognition approaches to recognize different tricks accurately. For uni-modal methods, we separately apply (1) CNN and LSTM; (2) CNN and BiLSTM; (3) CNN and BiLSTM with effective attention mechanisms; (4) Transformer-based action recognition pipeline. Transferred to the multi-modal conditions, we investigated the two-stream Inflated-3D architecture on "SkateboardAI" datasets to compare its performance with uni-modal cases. In sum, our objective is developing an excellent AI sport referee for the coolest skateboarding competitions.
Accelerating Human Action Recognition (HAR) efficiently for real-time surveillance and robotic systems on edge chips remains a challenging research field, given its high computational and memory requirements. This paper proposed an integrated end-to-end HAR scalable HW/SW accelerator co-design based on an enhanced 8-bit quantized Two-Stream SimpleNet-PyTorch CNN architecture. Our network accelerator was trained on UCF101 and UCF24 datasets and implemented on edge SoC-FPGA. Our development uses partially streaming dataflow architecture to achieve higher throughput versus network design and resource utilization trade-off. We also fused all convolutional, batch-norm, and ReLU operations into a single homogeneous layer and utilized the Lucas-Kanade motion flow method to enable a high parallelism accelerator design and optimized on-chip engine computing.Furthermore, our proposed methodology achieved nearly 81% prediction accuracy with an approximately 24 FPS real-time inference throughput at 187MHz on ZCU104, which is 1.7x - 1.9x higher than the prior research. Lastly, the designed framework was benchmarked against several hardware chips for higher throughput and performance measurements and is now available as an open-source project on GitHub for training and implementation on edge platforms.