EHTDI

Official implementation of "Exploring High-quality Target Domain Information for Unsupervised Domain Adaptive Semantic Segmentation". here

Accepted by ACM MM'22.

Our method does not require the distillation technique and requires only 1*3090 GPU.

News!!!!!!!!!!!!

We released the code for edge enhancement loss

Main Results

GTA5-to-CityScapes and SYNTHIA-to-CityScapes

	GTA5-to-CityScapes		SYNTHIA-to-CityScapes
	mIoU		mIoU_13 (mIoU_16)
Ours	58.8	Model	64.6 (57.8)	Model
Ours*	62.0	Model	69.2 (61.3)	Model

*Indicates a new edge enhancement loss is added and still no distillation technology is required.

Data Preparation

To run on GTA5-to-Cityscapes and SYNTHIA-to-Cityscapes, you need to download the respective datasets. Once they are downloaded, you can either modify the config files directly, or organize/symlink the data in the datasets/ directory as follows:

datasets
├── cityscapes
│   ├── gtFine
│   │   ├── train
│   │   │   ├── aachen
│   │   │   └── ...
│   │   └── val
│   └── leftImg8bit
│       ├── train
│       └── val
├── GTA5
│   ├── images
│   ├── labels
│   └── list
├── SYNTHIA
│   └── RAND_CITYSCAPES
│       ├── Depth
│       │   └── Depth
│       ├── GT
│       │   ├── COLOR
│       │   └── LABELS
│       ├── RGB
│       └── synthia_mapped_to_cityscapes
├── city_list
├── gta5_list
└── synthia_list

create symbolic link：

ln -s /data/lijj/pixmatch_output_61.5/ ./outputs

environment

requirement.txt

Initial Models

• ImageNet pretrain: Download
• For GTA5-to-Cityscapes, we start with a model pretrained on the source (GTA5): Download
• For SYNTHIA-to-Cityscapes, we start with a model pretrained on the source (SYNTHIA): Download

training

sh train.sh

Acknowledgments

This code is based on the implementations of PixMatch: Unsupervised Domain Adaptation via Pixelwise Consistency Training and DACS: Domain Adaptation via Cross-domain Mixed Sampling.