跳转至

文章背景与核心概要

传统的基于面部视频的自动应激检测方法通常在分类前将长时间的录像分割为短时间窗口。然而，这种切片方法带来了窗口长度、重叠率以及特征聚合等难题，同时也阻碍了模型直接捕捉整个录像中长距离时间动态的能力。

为了突破这些限制，本文作者推出了 FUSE（Frame-Unified Stress Estimation，帧统一应激估计） 这一创新框架。该框架能够将完整的视频录像作为单一的统一输入进行处理，彻底摆脱了时间窗口划分或外部切片的依赖。通过将时间维度折叠到空间表示的通道维度中，FUSE 实现了长视频的高效端到端评估，并在公开数据集上展现出了极具竞争力的实证性能，证明了在统一架构中进行完整录像推理的可行性。

FUSE: Frame-Unified Stress Estimation from Facial Video

FUSE: Frame-Unified Stress Estimation from Facial Video

Summary

Summary

Automatic stress detection from facial videos offers a non-intrusive approach to affect monitoring. However, traditional video-based methods typically segment long recordings into short temporal windows before classification. This approach introduces challenges regarding window length, overlap, and aggregation, while preventing the model from directly capturing long-range temporal dynamics across the entire recording.

Automatic stress detection from facial videos offers a non-intrusive approach to affect monitoring. However, traditional video-based methods typically segment long recordings into short temporal windows before classification. This approach introduces challenges regarding window length, overlap, and aggregation, while preventing the model from directly capturing long-range temporal dynamics across the entire recording.

To overcome these limitations, the authors introduce FUSE (Frame-Unified Stress Estimation), a novel framework that processes complete video recordings as a single unified input without temporal windowing or external segmentation.

To overcome these limitations, the authors introduce FUSE (Frame-Unified Stress Estimation), a novel framework that processes complete video recordings as a single unified input without temporal windowing or external segmentation.

Key Innovations and Findings:

  • Frame Unification: FUSE folds the temporal dimension into the channel dimension of the spatial representation, turning all frames into a single unified 2D tensor.
  • Unified Architecture: The high-dimensional input is processed using a specialized asymmetric-attention architecture. For example, a 120-second recording at 30 fps (3,600 frames) can be preserved and evaluated as a single input stream.
  • Empirical Performance: Evaluated on a 58-subject stress dataset using a stratified subject-level protocol across seven temporal-stride configurations:
  • Achieved a peak test accuracy of 69.44% at a temporal stride of \(t = 15\).
  • The full-frame configuration (\(t = 1\)) remained highly competitive, achieving 69.03% accuracy.
  • Computational costs ranged from 12.48 to 348.78 GFLOPs, illustrating the direct trade-off between temporal density and computational efficiency.

Key Innovations and Findings:

  • Frame Unification: FUSE folds the temporal dimension into the channel dimension of the spatial representation, turning all frames into a single unified 2D tensor.
  • Unified Architecture: The high-dimensional input is processed using a specialized asymmetric-attention architecture. For example, a 120-second recording at 30 fps (3,600 frames) can be preserved and evaluated as a single input stream.
  • Empirical Performance: Evaluated on a 58-subject stress dataset using a stratified subject-level protocol across seven temporal-stride configurations:
  • Achieved a peak test accuracy of 69.44% at a temporal stride of \(t = 15\).
  • The full-frame configuration (\(t = 1\)) remained highly competitive, achieving 69.03% accuracy.
  • Computational costs ranged from 12.48 to 348.78 GFLOPs, illustrating the direct trade-off between temporal density and computational efficiency.

These results demonstrate that temporal windowing is unnecessary for effective facial-video stress detection, proving that complete-recording inference is viable within a unified architecture.

These results demonstrate that temporal windowing is unnecessary for effective facial-video stress detection, proving that complete-recording inference is viable within a unified architecture.

Metadata

Metadata

  • arXiv ID: arXiv:2608.10442 [cs.CV]
  • DOI: 10.48550/arXiv.2608.10442
  • Primary Subject: Computer Vision and Pattern Recognition (cs.CV)
  • Secondary Subjects: Artificial Intelligence (cs.AI)
  • Authors:
  • Stefanos Gkikas
  • Thomas Kassiotis
  • Yang Guo
  • Guangliang Li
  • Giorgos Giannakakis
  • Submission History:
  • [v1] Tue, 11 Aug 2026
  • [v2] Tue, 18 Aug 2026 (Latest revision)
  • arXiv ID: arXiv:2608.10442 [cs.CV]
  • DOI: 10.48550/arXiv.2608.10442
  • Primary Subject: Computer Vision and Pattern Recognition (cs.CV)
  • Secondary Subjects: Artificial Intelligence (cs.AI)
  • Authors:
  • Stefanos Gkikas
  • Thomas Kassiotis
  • Yang Guo
  • Guangliang Li
  • Giorgos Giannakakis
  • Submission History:
  • [v1] Tue, 11 Aug 2026
  • [v2] Tue, 18 Aug 2026 (Latest revision)