Adaptation of Surgical Activity Recognition Models Across Operating Rooms

Automatic surgical activity recognition enables more intelligent devices and a efficient workflow. Integration of such technology in new operating rooms has the potential to improve care delivery patients decrease costs. Recent works have achieved promising performance on recognition; however, lack generalizability these models is one critical barriers wide-scale adoption this technology. In work, we study across rooms. We propose domain adaptation method model room for which only unlabeled videos. Our approach generates pseudo labels video clips that it confident about trains augmented version clips. extend our semi-supervised setting where small portion target also labeled. experiments, proposed consistently outperforms baselines dataset than 480 long videos collected from two