xmed-lab / distillingself Goto Github PK

Thanks for the code.
When I run the code according to README, the model is semi-supervised or fully supervised? If it is semi-supervised, how does the code choose labeled data from video 1-41?

Hi, thanks for sharing this project! This work is implemented with MS-TCN, does it use the information of future frames to predict the current frame? This may be a little different from other surgery workflow papers.

Another problem is that I found the reported performance (Acc=87.3, precision=73.5 and recall=85.1) on Cholec80 benchmark is much lower than other recent papers, like TMRNet (Acc=89.2, precision=89.7 and recall=89.5) and Trans-SVNet (Acc=90.3, precision=90.7 and recall=88.8). But the Jaccard performance is reverse, as DistillingSelf (jaccard=86.1) >> Trans-SVNet (79.3) > TMRNet (78.9). These results are confusing, and what causes the jaccard score achieve 7% advantage with inferior acc, precision and recall?

TMRNet: TMI-2021, https://github.com/YuemingJin/TMRNet
Trans-SVNet: MICCAI-2021, https://github.com/xjgaocs/Trans-SVNet