Requested for detailed downstream task preprocess code

[tianmaxingkong12]

Hi, thank you for your excellent work on OVFM and for releasing the code and pretrained weights.

I have a question regarding the downstream task fine-tuning pipeline. From my understanding, the released datasets are organized as original full-length surgical videos, while the downstream preprocess code(generate_video_paths_and_label) seems to use temporally sliced clips/windows as model inputs.

Could you kindly provide the preprocessing or data handling code for downstream fine-tuning, especially the part related to:

• converting full-length videos into temporal clips/windows,
• frame sampling strategy,
• clip-label alignment,
• and online/sliding-window input construction?

This would be very helpful for reproducing the downstream experiments and understanding the temporal training setup more clearly.

Thank you again for your great work and open-source contribution.