Title: Parameter-Efficient Image-to-Video Transfer for Fine-Grained Action Recognition

Abstract: This talk explores parameter-efficient image-to-video transfer for fine-grained action understanding. We examine parameter-efficient image-to-video adapters through three perspectives: 1) fine-grained temporal order of video frames, 2) motion frequency modeling, and 3) layer-wise adapter placement across transformer layers. These studies provide encouraging evidence that parameter-efficient image-to-video
adaptation can match or even outperform fully fine-tuned video models and existing PEFT approaches, while using substantially fewer trainable parameters.