Not End-to-End: Explore Multi-Stage Architecture for Online Surgical Phase Recognition

Surgical phase recognition is of particular interest to computer assisted surgery systems, in which the goal predict what occurring at each frame for a video. Networks with multi-stage architecture have been widely applied many vision tasks rich patterns, where predictor stage first outputs initial predictions and an additional refinement operates on perform further refinement. Existing works show that surgical video contents are well ordered contain temporal making suited task. However, we observe when simply applying task, end-to-end training manner will make ability fall short its wishes. To address problem, propose new non strategy explore different designs For strategy, trained separately proposed two types disturbed sequences. Meanwhile, evaluate three choices models our analysis solution robust specific models. We conduct experiments public benchmarks, M2CAI16 Workflow Challenge Cholec80 dataset. The SOTA comparable results holds great potential boost performance existing single-stage Code available https://github.com/ChinaYi/NETE.