DeepSceneSeg / EfficientPS

PyTorch code for training EfficientPS for Panoptic Segmentation

Home Page:http://panoptic.cs.uni-freiburg.de/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Intermediate logits's contribution?

CuberrChen opened this issue · comments

hi! The design of the network is very clever and it has achieved good results! But what I don't understand is why intermediate logits in fusion modeul are needed. According to my understanding,argmax(intermediate logits) actually represents that the possible instance part(In addition to the fusion results) of the semantic logits are accepted.
Hope to get your reply!
good luck.

The additional step is required to work around the sensitivity of the PQ metric by removing stuff classes that have very few pixels otherwise those pixels will be treated as false positives. The additional step improves the PQ metric a bit with respect to stuff classes.