Scholarlink

<p>While significant research efforts have been devoted to Weakly Supervised Video Anomaly Detection (WSVAD), prevailing approaches typically employ a two-stage paradigm: extracting frame-level features via visual encoders followed by temporal modeling on pre-computed features. Although effective, this decoupled framework inherently suffers from information degradation during feature extraction and suboptimal spatiotemporal fusion. To address these limitations, we propose a novel spatiotemporal adaptive framework that integrates temporal, spatial, and fusion adapters into the CLIP image encoder, enabling simultaneous spatiotemporal feature learning at the encoder level. Furthermore, we design a subspace alignment loss to ensure consistent embedding compatibility between the enhanced visual features and the original CLIP text space.Through comprehensive experimental validation on the UCF-Crime benchmark dataset, our framework demonstrates superior efficacy.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

TSAVD-CLIP: One visual encoder learn both temporal and spatial feat

  • 期刊论文

  • 发表日期：

    2025年7月26日

查看全文
  • Xu Jia,
  • Bin Chen,
  • Huaxiao Zhou,
  • Xiaochu Chen,
  • Lingxin Yu,
  • Zilun Zhang,
  • Yuxiang Cai

摘要

While significant research efforts have been devoted to Weakly Supervised Video Anomaly Detection (WSVAD), prevailing approaches typically employ a two-stage paradigm: extracting frame-level features via visual encoders followed by temporal modeling on pre-computed features. Although effective, this decoupled framework inherently suffers from information degradation during feature extraction and suboptimal spatiotemporal fusion. To address these limitations, we propose a novel spatiotemporal adaptive framework that integrates temporal, spatial, and fusion adapters into the CLIP image encoder, enabling simultaneous spatiotemporal feature learning at the encoder level. Furthermore, we design a subspace alignment loss to ensure consistent embedding compatibility between the enhanced visual features and the original CLIP text space.Through comprehensive experimental validation on the UCF-Crime benchmark dataset, our framework demonstrates superior efficacy.