EVENT DOSSIER
The researchers proposed a unsupervised test-time adaptation method called Cache Personalization (TTA-CaP), aimed at addressing the model adaptation issues in video facial expression recognition. This method utilizes three complementary mechanisms: personalized static caches, positive target caches, and negative target caches, combined with a three-stage control to prevent cache pollution and provide robust evidence. At the same time, fusion embedding is used to support stable video-level prediction. Experiments on the BioVid, StressID, and BAH datasets showed that TTA-CaP outperforms state-of-the-art methods in scenarios with subject-specific and environmental distribution shifts, while maintaining low computational and memory overhead. The related code is available on the GitHub repository.
本文提出一种名为 Cache Personalization (TTA-CaP) 的无监督测试时间适应方法，用于视频面部表情识别中通过缓存个性化实现低成本模型适配。该方法利用三个互补缓存（个性化静态缓存、正目标缓存和负目标缓存）及三门控机制来防止缓存污染并提供鲁棒性证据，同时结合嵌入融合支持稳定的视频级预测。在 BioVid、StressID 和 BAH 数据集上的实验表明，TTA-CaP 在主体特定和环境分布偏移下优于最先进的方法，且保持了低计算与内存开销。相关代码已公开于 GitHub 仓库。