Expanding Language-Image Pretrained Models for General Video Recognition

• Published in: Computer Vision – ECCV 2022 pp. 1 - 18
• Main Authors: Ni, Bolin, Peng, Houwen, Chen, Minghao, Zhang, Songyang, Meng, Gaofeng, Fu, Jianlong, Xiang, Shiming, Ling, Haibin
• Format: Book Chapter
• Language: English
• Published: Cham Springer Nature Switzerland 28.10.2022
• Series: Lecture Notes in Computer Science
• Subjects: Contrastive language-image pretraining; Video recognition
• DOI: 10.1007/978-3-031-19772-7_1
• ISBN: 9783031197710; 3031197712
• ISSN: 0302-9743

Contrastive language-image pretraining has shown great success in learning visual-textual joint representation from web-scale data, demonstrating remarkable “zero-shot” generalization ability for various image tasks. However, how to effectively expand such new language-image pretraining methods to video domains is still an open problem. In this work, we present a simple yet effective approach that adapts the pretrained language-image models to video recognition directly, instead of pretraining a new model from scratch. More concretely, to capture the long-range dependencies of frames along the temporal dimension, we propose a cross-frame attention mechanism that explicitly exchanges information across frames. Such module is lightweight and can be plugged into pretrained language-image models seamlessly. Moreover, we propose a video-specific prompting scheme, which leverages video content information for generating discriminative textual prompts. Extensive experiments demonstrate that our approach is effective and can be generalized to different video recognition scenarios. In particular, under fully-supervised settings, our approach achieves a top-1 accuracy of 87.1% on Kinectics-400, while using 12× $$\times $$ fewer FLOPs compared with Swin-L and ViViT-H. In zero-shot experiments, our approach surpasses the current state-of-the-art methods by +7.6% and +14.9% in terms of top-1 accuracy under two popular protocols. In few-shot scenarios, our approach outperforms previous best methods by +32.1% and +23.1% when the labeled data is extremely limited. Code and models are available at here.