search

fcdl94 / CoMFormer

Official implementation of "CoMFormer: Continual Learning in Semantic and Panoptic Segmentation"
https://arxiv.org/abs/2211.13999
Other
36 stars 3 forks source link

About Evaluate on VOC Pascal 2012 dataset #3

Closed XihuaQiao closed 10 months ago

XihuaQiao commented 10 months ago

Hi, thanks for your great work! I noticed the register_voc_segm.py file in your GitHub. Have you ever tested the CoMFormer in voc datasets? How is the performance? Cause I have been using Mask2Former on voc dataset and the results are disappointing.

fcdl94 commented 10 months ago

Hey! Yes, I tried with Pascal-VOC but I found it a bit unstable. I got good results, but I had to change them a bit the setting - such as removing an explicit background class and making it implicit. Anyway, I don't remember all the details... it was a bit ago

XihuaQiao commented 10 months ago

OMG, thank you for replying so quickly! I have tried the Mask2Former with the incremental setting as 16-5, and the base result of VOC in semantic segmentation is really low, with mIoU-Base = 74. I think it might be because the proportion of the background class is too large and MaskFormer architecture is not effective in this kind of situation.

fcdl94 commented 10 months ago

Take a look in some previous commits, I had some scripts on Pascal VOC and some way to train them!

XihuaQiao commented 10 months ago

Thanks a lot~