Questions about Semi-VOS

[9p15p]

Hi, Tanveer,

I have two questions.

1. When I evaluate or train the model. It has not use any reference information(e.g. mask or box). How does the model know what to segment for Semi-VOS.
   

   BoxVOS/adet/modeling/condinst/condinst.py

   Line 194 in 521e534

   features = self.backbone(images_norm.tensor)

2. Are there any codes for user to create the results for evaluation on DAVIS-2017(semi) or YoutubeVOS(semi). It needs to assigning colors according to reference information. The demo.py doesn't seem to have that function.

[Tanveer81]

The model only predicts objects per frame. We use PRemVOS to create the results for evaluation. We employ both optical flow and object re-identification as PReMVOS. We needed to adapt the code base of PReMVOS but did not change any underlying design. We will add the modified code and instructions soon.

[9p15p]

Hi, Tanveer, Could you please offer the Precomputed Results. Thank you.