A stupid question about auxiliary loss for Object detection & instance segmentation.

[waxnkw]

The paper said, "we do not have ground truth segmentation label.". I can understand there is no segmentation mask for detection, but why is there no segmentation mask for instance segmentation task?

[raoyongming]

Hi @waxnkw, thanks for your interest in our work. Here we want to say we don't have ground truth segmentation labels like semantic segmentation. For COCO, we cannot assign labels to the background regions if we only use the instance segmentation labels.

[waxnkw]

Thanks, so is the BCE loss applied to distinguish foreground objects and background?

[raoyongming]

We generate the segmentation mask using the bounding boxes or instance mask labels. We assign the corresponding labels for the foreground objects and set the rest regions as the background.

[waxnkw]

Thanks!