Adaptive Data Optimization for Cross-domain Action Recognition in Low-Light Environment

Human action recognition (HAR) serves as a foundational task in human-robot interaction applications. However, the challenge intensifies in low-light conditions, where limited dark scene training data are available, making current approaches less effective. In this paper, we propose a cross-domain adaptive data optimization approach to enhance human action recognition in dark lighting conditions. This approach encompasses two core components. The first one is adaptive data enhancement including stretching and zooming optical property transformation at the frame level. The second one is hybrid data filtering which progressively trains the model on data that exhibit a gradation of differences, ranging from minimal to substantial. We perform experiments of the comparison of our method and other baseline methods on the something-something v2 dataset and a small amount of dark scene action recognition data. The experiments demonstrate the effectiveness of our modules and and convenience of the overall pipeline.