Task 1. details

Hi, I have a question for the first task.

This category will require the teams to train multi-label fully supervised models. The model should classify all tools present within each frame of the video clips in the test set by training on the tool presence labels provided in the training set.

According to the explanation of the first task, the task is conducted within fully supervised setting. However, in the training set, only a ground-truth label for the entire video is provided. On the other hand, the task is to predict the tool presence label for "each frame". To clarify my understanding, the task will be weakly-supervised multi-label classification problem (the label is given only for the entire video.)?

Thank you in advance!

Hi ryo-hachiuma,

Yes, the predictions for the first category need to be tools present per frame (fully supervised classification). Though each video has one set of tool labels, the whole video will only have those tools present hence you can consider those labels as per frame (but some labels can be noisy - check out https://surgtoolloc.grand-challenge.org/data/).

The weakly supervised problem is primarily category 2 where teams need to only use tool presence labels for training and predict bounding boxes for each tool present.

Hope this answers your questions! Please feel free to reach out with any other questions.

Best, SurgToolLoc Organizing Committee

Hello,

I have a follow-up question, is the test labeled labeled the same way as the train set with the same combination of labels per video? Or is it labeled per frame where each video can have different combination of labels?

Kind regards,

Hi Aminey,

Thanks for your question. The test set will have frame wise labels, hence each video could have different set of lablels. Teams will be required to produce per frame predictions for both categories.

Best, SurgToolLoc 2022 Organizing Committee