Difference between arxiv paper v1 and the camera ready version? about k-net HOT 1 CLOSED

Hi @yucornetto ,
Thanks for your attention! After the paper release, we made some modifications of K-Net on panoptic segmentation from the original implementation as below:

1. In the arxiv v1 version, we use similar annotations as Panoptic FPN, which uses instance segmentation annotations for thing segmentation and use masks generated from panoptic annotations for stuff segmentation. However, when we are implementing Panoptic FPN in MMDet, we find that using panoptic annotations for both things and stuff segmentation yields better results (~1 PQ better). This is because the occlusion is not well handled between masks in the instance segmentation annotations. Therefore, we change to only use panoptic annotation in the new K-Net based on the newest version of MMDetection. This improves ~1.4 PQ of K-Net.
2. We fixed a bug in inference. We find a class ID bug when pasting things and stuff masks in arxiv v1 and fixed it when we change to use the COCOPanopticDataset in the newest MMDetection. Simply fixing this issue could improve about 1.0 PQ.
3. We tried a new inference strategy. In arxiv v1 we paste masks of things and stuff separately following the same way as Panoptic FPN. In rebuttal, we tried a new strategy that pastes them in a mixed order based on their classification scores. It consistently improves K-Net by 0.3~0.4 PQ.

from k-net.