The brain can only be fully understood through the lens of the behavior it generates -- a guiding principle in modern neuroscience research that nevertheless presents significant technical challenges. Many studies capture behavior with cameras, but video analysis approaches typically rely on specialized models requiring extensive labeled data. We address this limitation with BEAST (BEhavioral Analysis via Self-supervised pretraining of Transformers), a novel and scalable framework that pretrains experiment-specific vision transformers for diverse neuro-behavior analyses. BEAST combines masked autoencoding with temporal contrastive learning to effectively leverage unlabeled video data. Through comprehensive evaluation across multiple species, we demonstrate improved performance in three critical neuro-behavioral tasks: extracting behavioral features that correlate with neural activity, and pose estimation and action segmentation in both the single- and multi-animal settings. Our method establishes a powerful and versatile backbone model that accelerates behavioral analysis in scenarios where labeled data remains scarce.
Understanding behavior requires datasets that capture humans while carrying out complex tasks. The kitchen is an excellent environment for assessing human motor and cognitive function, as many complex actions are naturally exhibited in kitchens from chopping to cleaning. Here, we introduce the EPFL-Smart-Kitchen-30 dataset, collected in a noninvasive motion capture platform inside a kitchen environment. Nine static RGB-D cameras, inertial measurement units (IMUs) and one head-mounted HoloLens~2 headset were used to capture 3D hand, body, and eye movements. The EPFL-Smart-Kitchen-30 dataset is a multi-view action dataset with synchronized exocentric, egocentric, depth, IMUs, eye gaze, body and hand kinematics spanning 29.7 hours of 16 subjects cooking four different recipes. Action sequences were densely annotated with 33.78 action segments per minute. Leveraging this multi-modal dataset, we propose four benchmarks to advance behavior understanding and modeling through 1) a vision-language benchmark, 2) a semantic text-to-motion generation benchmark, 3) a multi-modal action recognition benchmark, 4) a pose-based action segmentation benchmark. We expect the EPFL-Smart-Kitchen-30 dataset to pave the way for better methods as well as insights to understand the nature of ecologically-valid human behavior. Code and data are available at https://github.com/amathislab/EPFL-Smart-Kitchen
Action segmentation is a core challenge in high-level video understanding, aiming to partition untrimmed videos into segments and assign each a label from a predefined action set. Existing methods primarily address single-person activities with fixed action sequences, overlooking multi-person scenarios. In this work, we pioneer textual reference-guided human action segmentation in multi-person settings, where a textual description specifies the target person for segmentation. We introduce the first dataset for Referring Human Action Segmentation, i.e., RHAS133, built from 133 movies and annotated with 137 fine-grained actions with 33h video data, together with textual descriptions for this new task. Benchmarking existing action recognition methods on RHAS133 using VLM-based feature extractors reveals limited performance and poor aggregation of visual cues for the target person. To address this, we propose a holistic-partial aware Fourier-conditioned diffusion framework, i.e., HopaDIFF, leveraging a novel cross-input gate attentional xLSTM to enhance holistic-partial long-range reasoning and a novel Fourier condition to introduce more fine-grained control to improve the action segmentation generation. HopaDIFF achieves state-of-the-art results on RHAS133 in diverse evaluation settings. The code is available at https://github.com/KPeng9510/HopaDIFF.git.
The skeleton-based temporal action segmentation (STAS) aims to densely segment and classify human actions within lengthy untrimmed skeletal motion sequences. Current methods primarily rely on graph convolutional networks (GCNs) for intraframe spatial modeling and temporal convolutional networks (TCNs) for interframe temporal modeling to discern motion patterns. However, these approaches often overlook the distinctive nature of essential action elements across various actions, including engaged core body parts and key subactions. This oversight limits the ability to distinguish different actions within a given sequence. To address these limitations, the snippet-aware Transformer with multiple action element (ME-ST) is proposed to enhance the discrimination and segmentation among actions, which leverages intrasnippet attention along joints and sequences to identify core joints and key subactions at different scales. Specifically, in terms of the spatial domain, the intrasnippet cross-joint attention (CJA) module divides the sequence into distinct snippets and computes attention to establish intricate joint semantic relationships, emphasizing the identification of core motion joints. In terms of the temporal domain, in the encoder, the intrasnippet cross-frame attention (CFA) module segments the sequence in a blockwise expansion manner and establishes interframe relationships to highlight the most discriminative frames. In the decoder, clip-level representations at various temporal scales are initially generated through an hourglass-like sampling process, followed by the intrasnippet cross-scale attention (CSA) module to integrate the key clip information across different time scales. The performance evaluation on five public datasets demonstrates that ME-ST achieves state-of-the-art (SOTA) performance.
Self-supervised representation learning for point cloud videos remains a challenging problem with two key limitations: (1) existing methods rely on explicit knowledge to learn motion, resulting in suboptimal representations; (2) prior Masked AutoEncoder (MAE) frameworks struggle to bridge the gap between low-level geometry and high-level dynamics in 4D data. In this work, we propose a novel self-disentangled MAE for learning expressive, discriminative, and transferable 4D representations. To overcome the first limitation, we learn motion by aligning high-level semantics in the latent space \textit{without any explicit knowledge}. To tackle the second, we introduce a \textit{self-disentangled learning} strategy that incorporates the latent token with the geometry token within a shared decoder, effectively disentangling low-level geometry and high-level semantics. In addition to the reconstruction objective, we employ three alignment objectives to enhance temporal understanding, including frame-level motion and video-level global information. We show that our pre-trained encoder surprisingly discriminates spatio-temporal representation without further fine-tuning. Extensive experiments on MSR-Action3D, NTU-RGBD, HOI4D, NvGesture, and SHREC'17 demonstrate the superiority of our approach in both coarse-grained and fine-grained 4D downstream tasks. Notably, Uni4D improves action segmentation accuracy on HOI4D by $+3.8\%$.
Temporal action segmentation (TAS) has long been a key area of research in both robotics and computer vision. In robotics, algorithms have primarily focused on leveraging proprioceptive information to determine skill boundaries, with recent approaches in surgical robotics incorporating vision. In contrast, computer vision typically relies on exteroceptive sensors, such as cameras. Existing multimodal TAS models in robotics integrate feature fusion within the model, making it difficult to reuse learned features across different models. Meanwhile, pretrained vision-only feature extractors commonly used in computer vision struggle in scenarios with limited object visibility. In this work, we address these challenges by proposing M2R2, a multimodal feature extractor tailored for TAS, which combines information from both proprioceptive and exteroceptive sensors. We introduce a novel pretraining strategy that enables the reuse of learned features across multiple TAS models. Our method achieves state-of-the-art performance on the REASSEMBLE dataset, a challenging multimodal robotic assembly dataset, outperforming existing robotic action segmentation models by 46.6%. Additionally, we conduct an extensive ablation study to evaluate the contribution of different modalities in robotic TAS tasks.
While there has been substantial progress in temporal action segmentation, the challenge to generalize to unseen views remains unaddressed. Hence, we define a protocol for unseen view action segmentation where camera views for evaluating the model are unavailable during training. This includes changing from top-frontal views to a side view or even more challenging from exocentric to egocentric views. Furthermore, we present an approach for temporal action segmentation that tackles this challenge. Our approach leverages a shared representation at both the sequence and segment levels to reduce the impact of view differences during training. We achieve this by introducing a sequence loss and an action loss, which together facilitate consistent video and action representations across different views. The evaluation on the Assembly101, IkeaASM, and EgoExoLearn datasets demonstrate significant improvements, with a 12.8% increase in F1@50 for unseen exocentric views and a substantial 54% improvement for unseen egocentric views.
Understanding a procedural activity requires modeling both how action steps transform the scene, and how evolving scene transformations can influence the sequence of action steps, even those that are accidental or erroneous. Existing work has studied procedure-aware video representations by proposing novel approaches such as modeling the temporal order of actions and has not explicitly learned the state changes (scene transformations). In this work, we study procedure-aware video representation learning by incorporating state-change descriptions generated by Large Language Models (LLMs) as supervision signals for video encoders. Moreover, we generate state-change counterfactuals that simulate hypothesized failure outcomes, allowing models to learn by imagining the unseen ``What if'' scenarios. This counterfactual reasoning facilitates the model's ability to understand the cause and effect of each step in an activity. To verify the procedure awareness of our model, we conduct extensive experiments on procedure-aware tasks, including temporal action segmentation and error detection. Our results demonstrate the effectiveness of the proposed state-change descriptions and their counterfactuals and achieve significant improvements on multiple tasks. We will make our source code and data publicly available soon.
Temporal action segmentation in untrimmed procedural videos aims to densely label frames into action classes. These videos inherently exhibit long-tailed distributions, where actions vary widely in frequency and duration. In temporal action segmentation approaches, we identified a bi-level learning bias. This bias encompasses (1) a class-level bias, stemming from class imbalance favoring head classes, and (2) a transition-level bias arising from variations in transitions, prioritizing commonly observed transitions. As a remedy, we introduce a constrained optimization problem to alleviate both biases. We define learning states for action classes and their associated transitions and integrate them into the optimization process. We propose a novel cost-sensitive loss function formulated as a weighted cross-entropy loss, with weights adaptively adjusted based on the learning state of actions and their transitions. Experiments on three challenging temporal segmentation benchmarks and various frameworks demonstrate the effectiveness of our approach, resulting in significant improvements in both per-class frame-wise and segment-wise performance.
This work presents the first condensation approach for procedural video datasets used in temporal action segmentation. We propose a condensation framework that leverages generative prior learned from the dataset and network inversion to condense data into compact latent codes with significant storage reduced across temporal and channel aspects. Orthogonally, we propose sampling diverse and representative action sequences to minimize video-wise redundancy. Our evaluation on standard benchmarks demonstrates consistent effectiveness in condensing TAS datasets and achieving competitive performances. Specifically, on the Breakfast dataset, our approach reduces storage by over 500$\times$ while retaining 83% of the performance compared to training with the full dataset. Furthermore, when applied to a downstream incremental learning task, it yields superior performance compared to the state-of-the-art.
Existing approaches to action segmentation use pre-computed frame features extracted by methods which have been trained on tasks that are different from action segmentation. Also, recent approaches typically use deep framewise representations that lack explicit modeling of action segments. To address these shortcomings, we introduce the first end-to-end solution to action segmentation -- End-to-End Action Segmentation Transformer (EAST). Our key contributions include: (1) a simple and efficient adapter design for effective backbone fine-tuning; (2) a segmentation-by-detection framework for leveraging action proposals initially predicted over a coarsely downsampled video toward labeling of all frames; and (3) a new action-proposal based data augmentation for robust training. EAST achieves state-of-the-art performance on standard benchmarks, including GTEA, 50Salads, Breakfast, and Assembly-101. The model and corresponding code will be released.
Existing supervised action segmentation methods depend on the quality of frame-wise classification using attention mechanisms or temporal convolutions to capture temporal dependencies. Even boundary detection-based methods primarily depend on the accuracy of an initial frame-wise classification, which can overlook precise identification of segments and boundaries in case of low-quality prediction. To address this problem, this paper proposes ASESM (Action Segmentation via Explicit Similarity Measurement) to enhance the segmentation accuracy by incorporating explicit similarity evaluation across frames and predictions. Our supervised learning architecture uses frame-level multi-resolution features as input to multiple Transformer encoders. The resulting multiple frame-wise predictions are used for similarity voting to obtain high quality initial prediction. We apply a newly proposed boundary correction algorithm that operates based on feature similarity between consecutive frames to adjust the boundary locations iteratively through the learning process. The corrected prediction is then further refined through multiple stages of temporal convolutions. As post-processing, we optionally apply boundary correction again followed by a segment smoothing method that removes outlier classes within segments using similarity measurement between consecutive predictions. Additionally, we propose a fully unsupervised boundary detection-correction algorithm that identifies segment boundaries based solely on feature similarity without any training. Experiments on 50Salads, GTEA, and Breakfast datasets show the effectiveness of both the supervised and unsupervised algorithms. Code and models are made available on Github.
Temporal logical understanding, a core facet of human cognition, plays a pivotal role in capturing complex sequential events and their temporal relationships within videos. This capability is particularly crucial in tasks like Video Question Answering (VideoQA), where the goal is to process visual data over time together with textual data to provide coherent answers. However, current VideoQA benchmarks devote little focus to evaluating this critical skill due to the challenge of annotating temporal logic. Despite the advancement of vision-language models, assessing their temporal logical reasoning powers remains a challenge, primarily due to the lack QA pairs that demand formal, complex temporal reasoning. To bridge this gap, we introduce the TimeLogic QA (TLQA) framework to automatically generate the QA pairs, specifically designed to evaluate the temporal logical understanding. To this end, TLQA leverages temporal annotations from existing video datasets together with temporal operators derived from logic theory to construct questions that test understanding of event sequences and their temporal relationships. TLQA framework is generic and scalable, capable of leveraging both, existing video action datasets with temporal action segmentation annotations, or video datasets with temporal scene graph annotations, to automatically generate temporal logical questions. We leverage 4 datasets, STAR, Breakfast, AGQA, and CrossTask, and generate two VideoQA dataset variants - small (TLQA-S) and large (TLQA-L) - containing 2k and 10k QA pairs for each category, resulting in 32k and 160k total pairs per dataset. We undertake a comprehensive evaluation of leading-edge VideoQA models, employing the TLQA to benchmark their temporal logical understanding capabilities. We assess the VideoQA model's temporal reasoning performance on 16 categories of temporal logic with varying temporal complexity.
This paper proposes a method for long-term action anticipation (LTA), the task of predicting action labels and their duration in a video given the observation of an initial untrimmed video interval. We build on an encoder-decoder architecture with parallel decoding and make two key contributions. First, we introduce a bi-directional action context regularizer module on the top of the decoder that ensures temporal context coherence in temporally adjacent segments. Second, we learn from classified segments a transition matrix that models the probability of transitioning from one action to another and the sequence is optimized globally over the full prediction interval. In addition, we use a specialized encoder for the task of action segmentation to increase the quality of the predictions in the observation interval at inference time, leading to a better understanding of the past. We validate our methods on four benchmark datasets for LTA, the EpicKitchen-55, EGTEA+, 50Salads and Breakfast demonstrating superior or comparable performance to state-of-the-art methods, including probabilistic models and also those based on Large Language Models, that assume trimmed video as input. The code will be released upon acceptance.
In this work, we address unsupervised temporal action segmentation, which segments a set of long, untrimmed videos into semantically meaningful segments that are consistent across videos. While recent approaches combine representation learning and clustering in a single step for this task, they do not cope with large variations within temporal segments of the same class. To address this limitation, we propose a novel method, termed Hierarchical Vector Quantization (HVQ), that consists of two subsequent vector quantization modules. This results in a hierarchical clustering where the additional subclusters cover the variations within a cluster. We demonstrate that our approach captures the distribution of segment lengths much better than the state of the art. To this end, we introduce a new metric based on the Jensen-Shannon Distance (JSD) for unsupervised temporal action segmentation. We evaluate our approach on three public datasets, namely Breakfast, YouTube Instructional and IKEA ASM. Our approach outperforms the state of the art in terms of F1 score, recall and JSD.
Existing skeleton-based human action classification models rely on well-trimmed action-specific skeleton videos for both training and testing, precluding their scalability to real-world applications where untrimmed videos exhibiting concatenated actions are predominant. To overcome this limitation, recently introduced skeleton action segmentation models involve un-trimmed skeleton videos into end-to-end training. The model is optimized to provide frame-wise predictions for any length of testing videos, simultaneously realizing action localization and classification. Yet, achieving such an improvement im-poses frame-wise annotated skeleton videos, which remains time-consuming in practice. This paper features a novel framework for skeleton-based action segmentation trained on short trimmed skeleton videos, but that can run on longer un-trimmed videos. The approach is implemented in three steps: Stitch, Contrast, and Segment. First, Stitch proposes a tem-poral skeleton stitching scheme that treats trimmed skeleton videos as elementary human motions that compose a semantic space and can be sampled to generate multi-action stitched se-quences. Contrast learns contrastive representations from stitched sequences with a novel discrimination pretext task that enables a skeleton encoder to learn meaningful action-temporal contexts to improve action segmentation. Finally, Segment relates the proposed method to action segmentation by learning a segmentation layer while handling particular da-ta availability. Experiments involve a trimmed source dataset and an untrimmed target dataset in an adaptation formulation for real-world skeleton-based human action segmentation to evaluate the effectiveness of the proposed method.
This paper presents a simple yet effective approach for the poorly investigated task of global action segmentation, aiming at grouping frames capturing the same action across videos of different activities. Unlike the case of videos depicting all the same activity, the temporal order of actions is not roughly shared among all videos, making the task even more challenging. We propose to use activity labels to learn, in a weakly-supervised fashion, action representations suitable for global action segmentation. For this purpose, we introduce a triadic learning approach for video pairs, to ensure intra-video action discrimination, as well as inter-video and inter-activity action association. For the backbone architecture, we use a Siamese network based on sparse transformers that takes as input video pairs and determine whether they belong to the same activity. The proposed approach is validated on two challenging benchmark datasets: Breakfast and YouTube Instructions, outperforming state-of-the-art methods.
Temporal action segmentation and long-term action anticipation are two popular vision tasks for the temporal analysis of actions in videos. Despite apparent relevance and potential complementarity, these two problems have been investigated as separate and distinct tasks. In this work, we tackle these two problems, action segmentation and action anticipation, jointly using a unified diffusion model dubbed ActFusion. The key idea to unification is to train the model to effectively handle both visible and invisible parts of the sequence in an integrated manner; the visible part is for temporal segmentation, and the invisible part is for future anticipation. To this end, we introduce a new anticipative masking strategy during training in which a late part of the video frames is masked as invisible, and learnable tokens replace these frames to learn to predict the invisible future. Experimental results demonstrate the bi-directional benefits between action segmentation and anticipation. ActFusion achieves the state-of-the-art performance across the standard benchmarks of 50 Salads, Breakfast, and GTEA, outperforming task-specific models in both of the two tasks with a single unified model through joint learning.
This paper introduces TemporalVLM, a video large language model capable of effective temporal reasoning and fine-grained understanding in long videos. At the core, our approach includes a visual encoder for mapping a long-term input video into features which are time-aware and contain both local and global cues. In particular, it first divides the input video into short-term clips, which are jointly encoded with their timestamps into time-sensitive local features. Next, the local features are passed through a bidirectional long short-term memory module for global feature aggregation. The extracted time-aware and multi-level features are important for accurate temporal reasoning and fine-grained understanding in long videos. Moreover, to facilitate the evaluation of TemporalVLM, we present a large-scale long video dataset of industry assembly processes, namely IndustryASM, which consists of videos recorded on factory floors with actions and timestamps annotated by industrial engineers for time and motion studies and temporal action segmentation evaluation. Finally, extensive experiments on datasets of long videos, including TimeIT and IndustryASM, show that TemporalVLM achieves superior performance than previous methods across temporal reasoning and fine-grained understanding tasks, namely dense video captioning, temporal video grounding, video highlight detection, and temporal action segmentation.
Temporal context plays a significant role in temporal action segmentation. In an offline setting, the context is typically captured by the segmentation network after observing the entire sequence. However, capturing and using such context information in an online setting remains an under-explored problem. This work presents the an online framework for temporal action segmentation. At the core of the framework is an adaptive memory designed to accommodate dynamic changes in context over time, alongside a feature augmentation module that enhances the frames with the memory. In addition, we propose a post-processing approach to mitigate the severe over-segmentation in the online setting. On three common segmentation benchmarks, our approach achieves state-of-the-art performance.