Pseudo-Augmentation Entropy Domain Adaptation for Video Unsupervised Action Recognition

This research tackles the challenge of video unsupervised domain adaptation (VUDA) in human action recognition, focusing on deploying models trained in well-lit conditions to poorly-lit environments. Current unsupervised domain adaptation (UDA) techniques mitigate performance degradation by aligning source and target domains through adversarial training. However, these approaches fail to enhance low-light images, resulting in information loss and suboptimal performance in dark settings. To address this issue, we introduce Pseudo-Augmentation Entropy Domain Adaptation (PANDA), an entropy theory-based method that simultaneously complements unsupervised adversarial alignment by incorporating image enhancement as an auxiliary task. Evaluations on the HMDB-ARID datasets demonstrate that proposed PANDA outperforms standard VUDA baselines, boosting action recognition accuracy by an average of 6.3%.