---

Iterative Loop Learning Combining Self-Training and Active Learning for Domain Adaptive Semantic Segmentation

by Licong Guan, Xue Yuan

This repository provides the official code for the paper Iterative Loop Learning Combining Self-Training and Active Learning for Domain Adaptive Semantic Segmentation.

Abstract Although data-driven methods have achieved great success in many tasks, it remains a significant challenge to ensure good generalization for different domain scenarios. Recently, self-training and active learning have been proposed to alleviate this problem. Self-training can improve model accuracy with massive unlabeled data, but some pseudo labels containing noise would be generated with limited or imbalanced training data. And there will be suboptimal models if human guidance is absent. Active learning can select more effective data to intervene, while the model accuracy can not be improved because the massive unlabeled data are not used. And the probability of querying sub-optimal samples will increase when the domain difference is too large, increasing annotation cost. This paper proposes an iterative loop learning method combining Self-Training and Active Learning (STAL) for domain adaptive semantic segmentation. The method first uses self-training to learn massive unlabeled data to improve model accuracy and provide more accurate selection models for active learning. Secondly, combined with the sample selection strategy of active learning, manual intervention is used to correct the self-training learning. Iterative loop to achieve the best performance with minimal label cost. Extensive experiments show that our method establishes state-of-the-art performance on tasks of GTAV→Cityscapes, SYNTHIA→Cityscapes, improving by 4.9% mIoU and 5.2% mIoU, compared to the previous best method, respectively.

For more information on STAL, please check our [Paper].

Usage

Prerequisites

• Python 3.6.9
• Pytorch 1.8.1
• torchvision 0.9.1

Step-by-step installation

git clone https://github.com/licongguan/STAL.git && cd STAL
conda create -n stal python=3.6.9
conda activate stal
pip install -r requirements.txt
pip install pip install torch==1.8.1+cu102 torchvision==0.9.1+cu102 -f https://download.pytorch.org/whl/torch_stable.html

Data Preparation

• Download The Cityscapes Dataset, The GTAV Dataset, and The SYNTHIA Dataset

For Cityscapes

Unzip the files to folder data and make the dictionary structures as follows:

data/dataset
├── gtFine
│   ├── test
│   ├── train
│   └── val
└── leftImg8bit
    ├── test
    ├── train
    └── val

For GTAV and SYNTHIA

Unzip the files to folder data and rename the image/label files for GTAV/SYNTHIA Datasets by running

python datasets/rename_gta5.py

next, move data to follow folder:

data/dataset
├── gtFine
│   ├── test
│   ├── train
│   │   ├── gta5
│   │   ├── synthia
│   └── val
└── leftImg8bit
    ├── test
    ├── train
    │   ├── gta5
    │   ├── synthia
    └── val

Prepare Pretrained Backbone

Before training, please download ResNet101 pretrained on ImageNet-1K from one of the following:

• Google Drive
• Baidu Drive Fetch Code: STAL

After that, modify model_urls in stal/models/resnet.py to </path/to/resnet101.pth>

Model Zoo

GTAV to Cityscapes

We have put our model checkpoints here [Google Drive] [百度网盘] (提取码STAL).

Method	Net	budget	mIoU	Chekpoint	Where in Our Paper
STAL	V3+	1%	70.0	Google Drive/BaiDu	Table1
STAL	V3+	2.2%	75.0	Google Drive/BaiDu	Table1
STAL	V3+	5.0%	76.1	Google Drive/BaiDu	Table1

SYNTHIA to Cityscapes

Method	Net	budget	mIoU	Chekpoint	Where in Our Paper
STAL	V3+	1%	73.2	Google Drive/BaiDu	Table2
STAL	V3+	2.2%	76.0	Google Drive/BaiDu	Table2
STAL	V3+	5.0%	76.6	Google Drive/BaiDu	Table2

STAL Training

We can train a model with labeled data from the GTAV dataset 2000 and labeled data from the Cityscapes dataset 30 (1%):

cd  experiments/splits/gtav2cityscapes/1.0%
# use torch.distributed.launch
sh train.sh <num_gpu> <port>

STAL Testing

sh eval.sh

Acknowledgement

This project is based on the following open-source projects: U²PL and RIPU. We thank their authors for making the source code publically available.

Citation

If you find this project useful in your research, please consider citing:

@inproceedings{guan2023iterative,
  title={Iterative Loop Learning Combining Self-Training and Active Learning for Domain Adaptive Semantic Segmentation},
  author={Guan, Licong and Yuan, Xue},
  journal={arXiv preprint arXiv:2301.13361},
  year={2023}
}

Contact

If you have any problem about our code, feel free to contact

• lcguan941@bjtu.edu.cn

or describe your problem in Issues.