cvlab-yonsei/dass

An official PyTorch implementation of "Bi-directional Contrastive Learning for Domain Adaptive Semantic Segmentation", ECCV 2022.

[ECCV22] Bi-directional Contrastive Learning for Domain Adaptive Semantic Segmentation

This is a PyTorch implementation of [Bi-directional Contrastive Learning for Domain adaptive Semantic Segmentation].

Requirements

To install requirements:

• Python 3.6
• Pytorch 1.4.0

Getting Started

1. Download the dataset.
2. Download the ImageNet-pretrained Model [Link].

Training

Train the source-only model:

python so_run.py

Train our model:

python run.py

Evaluation

To perform evaluation on single model:

python eval.py --frm model.pth --single

dass's Issues

Without finding so_config.yml file

Thanks for shiring the code, i can't find the so_config.yml file when using this code. Thank you

No configuration file

Dear author,
Thanks for sharing your code. I want to reproduce this work but I can not find the config file. Could you please release the config file?
Sincerely,
Junkun

Regarding EMA Update of Prototypes

Hi, interesting work. I couldn't seem to find the ema update of prototypes in the code as indicated in the paper. Am I missing something? Thanks.

Issues with dynamic pseudo labels

Dear DASS team,

I'm extremely interested in your work "Bi-directional Contrastive Learning for Domain Adaptive Semantic Segmentation", especially for the dynamic pseudo label part. It is an amazing idea to generate different pseudo-labels with different source prototypes so that denser pseudo-labels could be achieved and the harmful effect of few incorrect pseudo-labels could be mitigated. However, when I look into the repository, I don't find the code for dynamic pseudo-labels and hybrid pseudo-labels. Is this version of the repository up to date? If not could you please update the code?

Best,
Mingxuan