News!!! @ 2019/08/09.

We would like to support various backbones such as MobileNetV2, ResNet-101, Wide-ResNet-38, HRNetV2-48.

Update @ 2019/07/31.

We have released the paper ISA, which is very easy to use and implement while being much more efficient than the conventional self-attention based OCNet or DANet.

Update @ 2019/07/23.

We (HRNet + OCR w/ ASP) achieve Rank#1 on the leaderboard of Cityscapes (with a single model) on 3 of 4 metrics.

Update @ 2019/06/19.

We achieve 83.3116%+ on the leaderboard of Cityscapes test set based on single model HRNetV2 + OCR. Cityscapes leaderboard

We achieve 56.02% on the leaderboard of ADE20K test set based on single model ResNet101 + OCR without any bells or whistles. ADE20K leaderboard

Update @ 2019/05/27.

We achieve SOTA on 6 different semantic segmentation benchmarks including: Cityscapes, ADE20K, LIP, Pascal-Context, Pascal-VOC, COCO-Stuff. We provide the source code for our approach on all the six benchmarks. More benchmarks will be supported latter. We will consider release all the check-points and training log for the below experiments.

82.0%+/83.0%+ on the test set of Cityscapes with only Train-Fine + Val-Fine datasets/Coarse datasets.

45.5%+ on the val set of ADE20K.

56.5%+ on the val set of LIP.

56.0%+ on the val set of Pascal-Context.

81.0%+ on the val set of Pascal-VOC with ss test. (DeepLabv3+ is 80.02% with only train-aug)

40.5%+ on the val set of COCO-Stuff-10K.

• Cityscapes (testing with single scale whole image)

• ADE20K (testing with single scale whole image)

• LIP (testing with single scale whole image + left-right flip)

• Pascal-VOC (testing with single scale whole image)

• Pascal-Context (testing with single scale whole image)

• COCO-Stuff-10K (testing with single scale whole image)

Please consider citing our work if you find it helps you,

@article{yuan2018ocnet,
  title={Ocnet: Object context network for scene parsing},
  author={Yuan Yuhui and Wang Jingdong},
  journal={arXiv preprint arXiv:1809.00916},
  year={2018}
}

@article{huang2019isa,
  title={Interlaced Sparse Self-Attention for Semantic Segmentation},
  author={Huang Lang and Yuan Yuhui and Guo Jianyuan and Zhang Chao and Chen Xilin and Wang Jingdong},
  journal={arXiv preprint arXiv:1907.12273},
  year={2019}
}

This project is developed based on the segbox.pytorch and the author of segbox.pytorch donnyyou retains all the copyright of the reproduced Deeplabv3, PSPNet related code.