sdr

Introduction

Official Repo

Code Snippet

SDR (CVPR'2021)

@inproceedings{michieli2021continual,
  title={Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations},
  author={Michieli, Umberto and Zanuttigh, Pietro},
  booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},
  pages={1114--1124},
  year={2021}
}

Results

PASCAL VOC

Backbone	Crop Size	Setting	mIoU	Download
R-101-D16	512x512	15-5-disjoint		cfg | modellinks-per-step | log
R-101-D16	512x512	15-5-overlapped		cfg | modellinks-per-step | log
R-101-D16	512x512	15-1-disjoint		cfg | modellinks-per-step | log
R-101-D16	512x512	15-1-overlapped		cfg | modellinks-per-step | log
R-101-D16	512x512	10-1-disjoint		cfg | modellinks-per-step | log
R-101-D16	512x512	10-1-overlapped		cfg | modellinks-per-step | log

ADE20k

Backbone	Crop Size	Setting	mIoU	Download
R-101-D16	512x512	100-50-disjoint		cfg | modellinks-per-step | log
R-101-D16	512x512	100-50-overlapped		cfg | modellinks-per-step | log
R-101-D16	512x512	100-10-disjoint		cfg | modellinks-per-step | log
R-101-D16	512x512	100-10-overlapped		cfg | modellinks-per-step | log
R-101-D16	512x512	100-5-disjoint		cfg | modellinks-per-step | log
R-101-D16	512x512	100-5-overlapped		cfg | modellinks-per-step | log

More

You can also download the model weights from following sources:

• BaiduNetdisk: https://pan.baidu.com/s/1e9LlD6ITuLECstFcVPaCxQ with access code 55mq