Probing into Overfitting for Video Recognition

Yitian Zhang, Yue Bai, Huan Wang, Yizhou Wang, Yun Fu

Keywords: Action Recognition, Data Augmentation, Overfitting

TL;DR: We propose a data augmentation tailored for action recognition which shows consistent improvement over various models and datasets.

Abstract: Video recognition methods based on 2D networks have thrived in recent years, leveraging advanced image classification techniques. However, overfitting is an even severe problem in 2D video recognition models as 1) the scale of video datasets is relatively small compared to image recognition datasets like ImageNet; 2) current pipeline treats background and semantic frames equally during optimization which aggravates overfitting. Based on these challenges, we design a video-specific data augmentation approach, named as Ghost Motion (GM), to alleviate overfitting. Specifically, GM shifts channels along temporal dimension to enable semantic motion information diffused into other frames which may be irrelevant originally, leading to improvement in frame-wise accuracy. In addition, for challenging video samples with significant temporal dependency (e.g., Something-Something), we further scale the logits during training to prevent overconfident predictions on background frames. Comprehensive empirical validation on various popular datasets shows that the proposed method can improve the generalization of existing methods and is compatible to other competing data augmentation approaches.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)