Scholarlink

<p>Video understanding requires not only recognizing actions but also pinpointing the precise moments when objects undergo critical state transitions. Object State Changes (OSCs) aim to precisely locate state transitions within temporal sequences. Existing approaches often overlook the underlying state change-driven mechanisms, which limits their reasoning ability. Moreover, by modeling OSCs as hard-label classification tasks, they fail to capture inter-frame continuity, leading to unstable, non-smooth predictions. To overcome these challenges, we propose a variational causal inference-based reasoning method, VCI-OSC, which incorporates a Structural Causal Model (SCM) to model OSCs. It features two novel components: a Variational Reasoning Module (VRM) and a State Classifier (SC). The VRM models unobservable latent factors via a Gaussian encoder and reparametrized sampling, while being optimized with a combined reconstruction and KL divergence loss; SC incorporates learnable ordered weights to enable soft partitioning and continuous modeling of states, improving precision and generalization capability in recognizing state transitions. Experimental results on three benchmarks show that VCI-OSC significantly outperforms existing methods in both precision and generalization. This work not only advances the theoretical understanding of OSCs but also offers practical value for enhancing the performance of process monitoring systems in industrial applications.</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

A variational causal inference-based method for recognizing object state changes in videos

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Zhichao Wang,
  • Wenliang Ge,
  • Shucheng Huang,
  • Mingxing Li

摘要

Video understanding requires not only recognizing actions but also pinpointing the precise moments when objects undergo critical state transitions. Object State Changes (OSCs) aim to precisely locate state transitions within temporal sequences. Existing approaches often overlook the underlying state change-driven mechanisms, which limits their reasoning ability. Moreover, by modeling OSCs as hard-label classification tasks, they fail to capture inter-frame continuity, leading to unstable, non-smooth predictions. To overcome these challenges, we propose a variational causal inference-based reasoning method, VCI-OSC, which incorporates a Structural Causal Model (SCM) to model OSCs. It features two novel components: a Variational Reasoning Module (VRM) and a State Classifier (SC). The VRM models unobservable latent factors via a Gaussian encoder and reparametrized sampling, while being optimized with a combined reconstruction and KL divergence loss; SC incorporates learnable ordered weights to enable soft partitioning and continuous modeling of states, improving precision and generalization capability in recognizing state transitions. Experimental results on three benchmarks show that VCI-OSC significantly outperforms existing methods in both precision and generalization. This work not only advances the theoretical understanding of OSCs but also offers practical value for enhancing the performance of process monitoring systems in industrial applications.