Spatio-Temporal Joint Density Driven Learning for Skeleton-Based Action Recognition

Traditional approaches in unsupervised or self supervised learning for skeleton-based action classification have concentrated predominantly on the dynamic aspects of skeletal sequences. Yet, the intricate interaction between the moving and static elements of the skeleton presents a rarely tapped discriminative potential for action classification. This paper introduces a novel measurement, referred to as spatial-temporal joint density (STJD), to quantify such interaction. Tracking the evolution of this density throughout an action can effectively identify a subset of discriminative moving and/or static joints termed "prime joints" to steer self-supervised learning. A new contrastive learning strategy named STJD-CL is proposed to align the representation of a skeleton sequence with that of its prime joints while simultaneously contrasting the representations of prime and nonprime joints. In addition, a method called STJD-MP is developed by integrating it with a reconstruction-based framework for more effective learning. Experimental evaluations on the NTU RGB+D 60, NTU RGB+D 120, and PKUMMD datasets in various downstream tasks demonstrate that the proposed STJD-CL and STJD-MP improved performance, particularly by 3.5 and 3.6 percentage points over the state-of-the-art contrastive methods on the NTU RGB+D 120 dataset using X-sub and X-set evaluations, respectively.

Reward-Driven Interaction: Enhancing Proactive Dialogue Agents through User Satisfaction Prediction

Reward-driven proactive dialogue agents require precise estimation of user satisfaction as an intrinsic reward signal to determine optimal interaction strategies. Specifically, this framework triggers clarification questions when detecting potential user dissatisfaction during interactions in the industrial dialogue system. Traditional works typically rely on training a neural network model based on weak labels which are generated by a simple model trained on user actions after current turn. However, existing methods suffer from two critical limitations in real-world scenarios: (1) Noisy Reward Supervision, dependence on weak labels derived from post-hoc user actions introduces bias, particularly failing to capture satisfaction signals in ASR-error-induced utterances; (2) Long-Tail Feedback Sparsity, the power-law distribution of user queries causes reward prediction accuracy to drop in low-frequency domains. The noise in the weak labels and a power-law distribution of user utterances results in that the model is hard to learn good representation of user utterances and sessions. To address these limitations, we propose two auxiliary tasks to improve the representation learning of user utterances and sessions that enhance user satisfaction prediction. The first one is a contrastive self-supervised learning task, which helps the model learn the representation of rare user utterances and identify ASR errors. The second one is a domain-intent classification task, which aids the model in learning the representation of user sessions from long-tailed domains and improving the model's performance on such domains. The proposed method is evaluated on DuerOS, demonstrating significant improvements in the accuracy of error recognition on rare user utterances and long-tailed domains.

SEVERE++: Evaluating Benchmark Sensitivity in Generalization of Video Representation Learning

Continued advances in self-supervised learning have led to significant progress in video representation learning, offering a scalable alternative to supervised approaches by removing the need for manual annotations. Despite strong performance on standard action recognition benchmarks, video self-supervised learning methods are largely evaluated under narrow protocols, typically pretraining on Kinetics-400 and fine-tuning on similar datasets, limiting our understanding of their generalization in real world scenarios. In this work, we present a comprehensive evaluation of modern video self-supervised models, focusing on generalization across four key downstream factors: domain shift, sample efficiency, action granularity, and task diversity. Building on our prior work analyzing benchmark sensitivity in CNN-based contrastive learning, we extend the study to cover state-of-the-art transformer-based video-only and video-text models. Specifically, we benchmark 12 transformer-based methods (7 video-only, 5 video-text) and compare them to 10 CNN-based methods, totaling over 1100 experiments across 8 datasets and 7 downstream tasks. Our analysis shows that, despite architectural advances, transformer-based models remain sensitive to downstream conditions. No method generalizes consistently across all factors, video-only transformers perform better under domain shifts, CNNs outperform for fine-grained tasks, and video-text models often underperform despite large scale pretraining. We also find that recent transformer models do not consistently outperform earlier approaches. Our findings provide a detailed view of the strengths and limitations of current video SSL methods and offer a unified benchmark for evaluating generalization in video representation learning.

H-MoRe: Learning Human-centric Motion Representation for Action Analysis

In this paper, we propose H-MoRe, a novel pipeline for learning precise human-centric motion representation. Our approach dynamically preserves relevant human motion while filtering out background movement. Notably, unlike previous methods relying on fully supervised learning from synthetic data, H-MoRe learns directly from real-world scenarios in a self-supervised manner, incorporating both human pose and body shape information. Inspired by kinematics, H-MoRe represents absolute and relative movements of each body point in a matrix format that captures nuanced motion details, termed world-local flows. H-MoRe offers refined insights into human motion, which can be integrated seamlessly into various action-related applications. Experimental results demonstrate that H-MoRe brings substantial improvements across various downstream tasks, including gait recognition(CL@R1: +16.01%), action recognition(Acc@1: +8.92%), and video generation(FVD: -67.07%). Additionally, H-MoRe exhibits high inference efficiency (34 fps), making it suitable for most real-time scenarios. Models and code will be released upon publication.

Towards Micro-Action Recognition with Limited Annotations: An Asynchronous Pseudo Labeling and Training Approach

Micro-Action Recognition (MAR) aims to classify subtle human actions in video. However, annotating MAR datasets is particularly challenging due to the subtlety of actions. To this end, we introduce the setting of Semi-Supervised MAR (SSMAR), where only a part of samples are labeled. We first evaluate traditional Semi-Supervised Learning (SSL) methods to SSMAR and find that these methods tend to overfit on inaccurate pseudo-labels, leading to error accumulation and degraded performance. This issue primarily arises from the common practice of directly using the predictions of classifier as pseudo-labels to train the model. To solve this issue, we propose a novel framework, called Asynchronous Pseudo Labeling and Training (APLT), which explicitly separates the pseudo-labeling process from model training. Specifically, we introduce a semi-supervised clustering method during the offline pseudo-labeling phase to generate more accurate pseudo-labels. Moreover, a self-adaptive thresholding strategy is proposed to dynamically filter noisy labels of different classes. We then build a memory-based prototype classifier based on the filtered pseudo-labels, which is fixed and used to guide the subsequent model training phase. By alternating the two pseudo-labeling and model training phases in an asynchronous manner, the model can not only be learned with more accurate pseudo-labels but also avoid the overfitting issue. Experiments on three MAR datasets show that our APLT largely outperforms state-of-the-art SSL methods. For instance, APLT improves accuracy by 14.5\% over FixMatch on the MA-12 dataset when using only 50\% labeled data. Code will be publicly available.

Surg-3M: A Dataset and Foundation Model for Perception in Surgical Settings

Advancements in computer-assisted surgical procedures heavily rely on accurate visual data interpretation from camera systems used during surgeries. Traditional open-access datasets focusing on surgical procedures are often limited by their small size, typically consisting of fewer than 100 videos with less than 100K images. To address these constraints, a new dataset called Surg-3M has been compiled using a novel aggregation pipeline that collects high-resolution videos from online sources. Featuring an extensive collection of over 4K surgical videos and more than 3 million high-quality images from multiple procedure types, Surg-3M offers a comprehensive resource surpassing existing alternatives in size and scope, including two novel tasks. To demonstrate the effectiveness of this dataset, we present SurgFM, a self-supervised foundation model pretrained on Surg-3M that achieves impressive results in downstream tasks such as surgical phase recognition, action recognition, and tool presence detection. Combining key components from ConvNeXt, DINO, and an innovative augmented distillation method, SurgFM exhibits exceptional performance compared to specialist architectures across various benchmarks. Our experimental results show that SurgFM outperforms state-of-the-art models in multiple downstream tasks, including significant gains in surgical phase recognition (+8.9pp, +4.7pp, and +3.9pp of Jaccard in AutoLaparo, M2CAI16, and Cholec80), action recognition (+3.1pp of mAP in CholecT50) and tool presence detection (+4.6pp of mAP in Cholec80). Moreover, even when using only half of the data, SurgFM outperforms state-of-the-art models in AutoLaparo and achieves state-of-the-art performance in Cholec80. Both Surg-3M and SurgFM have significant potential to accelerate progress towards developing autonomous robotic surgery systems.

JARVIS-VLA: Post-Training Large-Scale Vision Language Models to Play Visual Games with Keyboards and Mouse

Recently, action-based decision-making in open-world environments has gained significant attention. Visual Language Action (VLA) models, pretrained on large-scale web datasets, have shown promise in decision-making tasks. However, previous work has primarily focused on action post-training, often neglecting enhancements to the foundational model itself. In response, we introduce a novel approach, Act from Visual Language Post-Training, which refines Visual Language Models (VLMs) through visual and linguistic guidance in a self-supervised manner. This enhancement improves the models' capabilities in world knowledge, visual recognition, and spatial grounding in open-world environments. Following the above post-training paradigms, we obtain the first VLA models in Minecraft that can follow human instructions on over 1k different atomic tasks, including crafting, smelting, cooking, mining, and killing. Our experiments demonstrate that post-training on non-trajectory tasks leads to a significant 40% improvement over the best agent baseline on a diverse set of atomic tasks. Furthermore, we demonstrate that our approach surpasses traditional imitation learning-based policies in Minecraft, achieving state-of-the-art performance. We have open-sourced the code, models, and datasets to foster further research. The project page can be found in https://craftjarvis.github.io/JarvisVLA.

Deep Learning For Time Series Analysis With Application On Human Motion

Time series data, defined by equally spaced points over time, is essential in fields like medicine, telecommunications, and energy. Analyzing it involves tasks such as classification, clustering, prototyping, and regression. Classification identifies normal vs. abnormal movements in skeleton-based motion sequences, clustering detects stock market behavior patterns, prototyping expands physical therapy datasets, and regression predicts patient recovery. Deep learning has recently gained traction in time series analysis due to its success in other domains. This thesis leverages deep learning to enhance classification with feature engineering, introduce foundation models, and develop a compact yet state-of-the-art architecture. We also address limited labeled data with self-supervised learning. Our contributions apply to real-world tasks, including human motion analysis for action recognition and rehabilitation. We introduce a generative model for human motion data, valuable for cinematic production and gaming. For prototyping, we propose a shape-based synthetic sample generation method to support regression models when data is scarce. Lastly, we critically evaluate discriminative and generative models, identifying limitations in current methodologies and advocating for a robust, standardized evaluation framework. Our experiments on public datasets provide novel insights and methodologies, advancing time series analysis with practical applications.

Rethinking Top Probability from Multi-view for Distracted Driver Behaviour Localization

Naturalistic driving action localization task aims to recognize and comprehend human behaviors and actions from video data captured during real-world driving scenarios. Previous studies have shown great action localization performance by applying a recognition model followed by probability-based post-processing. Nevertheless, the probabilities provided by the recognition model frequently contain confused information causing challenge for post-processing. In this work, we adopt an action recognition model based on self-supervise learning to detect distracted activities and give potential action probabilities. Subsequently, a constraint ensemble strategy takes advantages of multi-camera views to provide robust predictions. Finally, we introduce a conditional post-processing operation to locate distracted behaviours and action temporal boundaries precisely. Experimenting on test set A2, our method obtains the sixth position on the public leaderboard of track 3 of the 2024 AI City Challenge.

Spatial Hierarchy and Temporal Attention Guided Cross Masking for Self-supervised Skeleton-based Action Recognition

In self-supervised skeleton-based action recognition, the mask reconstruction paradigm is gaining interest in enhancing model refinement and robustness through effective masking. However, previous works primarily relied on a single masking criterion, resulting in the model overfitting specific features and overlooking other effective information. In this paper, we introduce a hierarchy and attention guided cross-masking framework (HA-CM) that applies masking to skeleton sequences from both spatial and temporal perspectives. Specifically, in spatial graphs, we utilize hyperbolic space to maintain joint distinctions and effectively preserve the hierarchical structure of high-dimensional skeletons, employing joint hierarchy as the masking criterion. In temporal flows, we substitute traditional distance metrics with the global attention of joints for masking, addressing the convergence of distances in high-dimensional space and the lack of a global perspective. Additionally, we incorporate cross-contrast loss based on the cross-masking framework into the loss function to enhance the model's learning of instance-level features. HA-CM shows efficiency and universality on three public large-scale datasets, NTU-60, NTU-120, and PKU-MMD. The source code of our HA-CM is available at https://github.com/YinxPeng/HA-CM-main.