search

HRNet / HRNet-Semantic-Segmentation

The OCR approach is rephrased as Segmentation Transformer: https://arxiv.org/abs/1909.11065. This is an official implementation of semantic segmentation for HRNet. https://arxiv.org/abs/1908.07919
Other
3.14k stars 686 forks source link

SegFix Operation ? #90

Open AssassinCroc opened 4 years ago

AssassinCroc commented 4 years ago

Firstly, congrats, this is great work and thank you for releasing code. I was wondering what SegFix is ? I saw that on cityscapes leaderboard, you mentioned SegFix.

-Thanks

PkuRainBow commented 4 years ago

SegFix is another work under review, which is a model-agnostic post-processing scheme that can address the boundary errors for both semantic segmentation and instance segmentation. We would like to release the paper and code of SegFix once published.

AssassinCroc commented 4 years ago

@PkuRainBow very interesting. Looking forward to it. One more question out of curiosity. Is this some sort of model or post processing algorithm ?

howard-mahe commented 4 years ago

Any update regarding SegFix ?

PkuRainBow commented 4 years ago

@howard-mahe We have released the testing files & scripts of SegFix here

howard-mahe commented 4 years ago

Thank you, I have not seen it before. Do you have any idea when do you plan to upload the preprint on arXiv?

PkuRainBow commented 4 years ago

@howard-mahe Thanks for your interest and we plan to release the paper after the ECCV decision. Please feel free to try our SegFix on the Cityscapes segmentation (semantic & instance) benchmarks, which is also the main point we want to sell.

davidblom603 commented 4 years ago

Cool that you've released the testing files & script of SegFix. Can you apply these scripts also easily to other datasets?

PkuRainBow commented 4 years ago

@davidblom603 Currently, the testing files & scripts can only be applied to fix the boundary errors on the Cityscapes dataset. We will also release the training scripts of SegFix once the paper is accepted.

davidblom603 commented 4 years ago

Ah oke. Thanks.