A Domain Confusion Adaptation Method for Semi-supervised Action Recognition in Dark Environments
Guanzhi Ding, Peng Gao, Haojie Wu · 2023
Action recognition in the dark (ARID) is a significant task, which is widely utilized in self-driving at night, night surveillance, etc. However, current fully supervised networks require additional annotation costs to generate large-scale labeled datasets in dark environments for training, which is unreasonable. A feasible solution is to achieve domain transfer between labeled normal illumination datasets and unlabeled datasets in dark environments. Therefore, we propose a novel two-stream CNN architecture employing a semi-supervised learning strategy to facilitate domain transfer. Specifically, the architecture introduces a brightness enhancement module to enhance feature representations in dark environments and an additional domain confusion loss to jointly align domain-invariant representations between labeled source domains and unlabeled target domains. Our domain adaptation method achieves superior performance on the semi-supervised ARID dataset.