search

HAL-42 / FMA-WSSS

This repo is a implementation of the Foundation Model Assisted Weakly Supervised Semantic Segmentation. The code is developed based on the Pytorch framework.
Apache License 2.0
38 stars 4 forks source link

mIoU without the SAM for the post processing in the V3+ and SwinL results. #31

Open HYTHYThythyt opened 2 weeks ago

HYTHYThythyt commented 2 weeks ago

Hi, author. What are the mIoUs in the final segmentation results that did not using the SAM for the post processing in V3+ and SwinL, respectively? And also I want to know, what are the mIoUs in final segmentation results that you using the CRF for the post processing in V3+ and SwinL, respectively? I found that the V3+ need to crop the truth-labels when evaluate the val and test set, there are very poor results when did not using the crop in val and test set, and also the test set did not be evaluated successfully when crop the images, because there are not crop version for the test set provided by the official server. But your results are very great both in V3+ and SwinL. Are there any different config between yours and the others? Any help is very appreciate! Thank you!

HAL-42 commented 2 weeks ago
  1. The final Swin-L results didn't use any post-processing.
  2. The no post-processing or CRF results are reported at Table.3 of our paper.
  3. Standard evaluation should not involve cropping ground truth. What is this "crop" mean, or who is "the others" that need this "crop" for evaluating?
HYTHYThythyt commented 2 weeks ago

The crop means cropping. The V3+ projects have the ‘crop’ options in some previous papers, I do not mean they certainly using the 'crop' options to cropping the ground-truth labels. Thank you for you patient reply.

HYTHYThythyt commented 4 days ago

Hi, author, I want to know which configs are for the deeplabv3+ to train the voc and coco, respectively. Thank you!!!

HYTHYThythyt commented 4 days ago

I had found many config files in the dir(deeplabv3plus), but I do not know which one is the final config for v3+ to train voc and coco, respectively. Any help is very appreciate! Thank you!

HAL-42 commented 3 days ago

VOC

mmsegmentation trianing config: others/mmsegmentation/configs/deeplabv3plus/d3p-r1d8-bt16-80e-512x-背6s-2au.py

SAMS post-processing config: configs/anns_seed/val/mm/ann=rsw3,prob=d3p-r1d8-bt16-80e-512x-背6s-2au,ms/cfg.py

COCO

mmsegmentation trianing config: others/mmsegmentation/configs/deeplabv3plus/d3p-r1d8-bt16-160e-512x-ES-2au.py

SAMS post-processing config: configs/anns_seed/coco/v/ann=rsw3,prob=d3p-r1d8-bt16-160e-512x-ES-2au,ms/cfg.py