Optimal Spatio-Temporal Descriptor for Generalizable Video Recognition
To address the semantic gap between web-scaled descriptive narratives and concise action category names, we propose to disentangle category names into Spatio-Temporal Descriptors using large language models. We further introduce Optimal Descriptor Solver to adaptively align frame-level representations with the refined textual knowledge, enabling generalizable video recognition.
Efficient Video Recognition in Long-untrimmed Videos: View while Moving
Proposing the "View while Moving" paradigm for efficient video recognition in long-untrimmed videos, accessing raw frames once during inference and achieving improved accuracy and efficiency trade-offs.
Percept, Chat, and Adapt: Multimodal Knowledge Transfer of Foundation Models for Open-World Video Recognition
Foundation models with rich knowledge can boost open-world video recognition through a generic knowledge transfer pipeline named PCA.
Unveiling the Benefits of Hue Variance in Video Recognition with Motion Coherent Augmentation
Hue variance benefits video recognition by prioritizing motion patterns over static appearances.