Question on model mask output

[chorongi]

Hello, I'm trying to train DeVIS with custom data. (Which contains only 1 class, i.e. a MOT like dataset) I observe that the output mask for each instance track contains the segmentation mask for all instances inside the frame instead of just one. (I was expecting each query would give a single instance-level segmentation instead of per-class or per-frame segmentation) Is this an expected behavior?? If you have experienced this before, any kind of advice would be appreciated!

Also, will the model checkpoints become available at any point?

[timmeinhardt]

We fixed the link for the model checkpoints. They should be downloadable now.

Regarding the instance masks for your single-class case, I assume this is a bug which occurs for this edge case. Probably in the post-processor. However, I was not able to find it immediately since @acaelles97 is more familiar with the codebase.