DVIS-DAQ is an impressive work. I have a question regarding the training strategy:

The paper mentions that the segmenter, tracker, and temporal refiner are trained in separate stages with frozen components. From the code, I see that MinVIS, Online, and Offline modes can be used to train these three components individually. However, I couldn’t find where to configure the corresponding weights for each stage.

Should I load the weights from the previous stage in the configuration file like this?

MODEL:
  WEIGHTS: 'MinVIS.pth'

MODEL:
  WEIGHTS: 'Online.pth'

Additionally, I noticed that you have shared pre-trained weights on Hugging Face. Could you let me know which one corresponds to the COCO & YTVIS-2021 VIT finetuned segmenter?