Multi-Anchor Active Domain Adaptation for Semantic Segmentation
Munan Ning*, Donghuan Lu*, Dong Wei†, Cheng Bian, Chenglang Yuan, Shuang Yu, Kai Ma, Yefeng Zheng
[Paper] [PPT] [Graphic Abstract]
Table of Contents
Introduction
This respository contains the MADA method as described in the ICCV 2021 Oral paper "Multi-Anchor Active Domain Adaptation for Semantic Segmentation".
Requirements
The code requires Pytorch >= 0.4.1 with python 3.6. The code is trained using a NVIDIA Tesla V100 with 32 GB memory. You can simply reduce the batch size in stage 2 to run on a smaller memory.
Usage
- Preparation:
- Download the GTA5 dataset as the source domain, and the Cityscapes dataset as the target domain.
- Download the weights and features. Move features to the MADA directory.
- Setup the config files.
- Set the data paths
- Set the pretrained model paths
- Training-quick
- To run the code with our weights and anchors (anchors/cluster_centroids_full_10.pkl):
python3 train_active_stage1.py
python3 train_active_stage2.py
- During the training, the generated files (log file) will be written in the folder 'runs/..'.
- Evaluation
- Set the config file for test (configs/test_from_city_to_gta.yml):
- Run:
python3 test.py
to see the results.
- Training-whole process
- Setting the config files.
- Stage 1:
- 1-save_feat_source.py: get the './features/full_dataset_objective_vectors.pkl'
python3 save_feat_source.py
- 2-cluster_anchors_source.py: cluster the './features/full_dataset_objective_vectors.pkl' to './anchors/cluster_centroids_full_10.pkl'
python3 cluster_anchors_source.py
- 3-select_active_samples.py: select active samples with './anchors/cluster_centroids_full_10.pkl' to 'stage1_cac_list_0.05.txt'
python3 select_active_samples.py
- 4-train_active_stage1.py: train stage1 model with anchors './anchors/cluster_centroids_full_10.pkl' and active samples 'stage1_cac_list_0.05.txt', get the 'from_gta5_to_cityscapes_on_deeplab101_best_model_stage1.pkl', which is stored in the runs/active_from_gta_to_city_stage1
python3 train_active_stage1.py
- Stage 2:
- 1-save_feat_target.py: get the './features/target_full_dataset_objective_vectors.pkl.pkl'
python3 save_feat_target.py
- 2-cluster_anchors_target.py: cluster the './features/target_full_dataset_objective_vectors.pkl' to './anchors/cluster_centroids_full_target_10.pkl'
python3 cluster_anchors_target.py
- 3-train_active_stage2.py: train stage2 model with anchors './anchors/cluster_centroids_full_target_10.pkl' and active samples 'stage1_cac_list_0.05.txt', get the 'from_gta5_to_cityscapes_on_deeplab101_best_model_stage2.pkl'
python3 train_active_stage2.py
License
The code is heavily borrowed from the CAG_UDA (https://github.com/RogerZhangzz/CAG_UDA).
If you use this code and find it usefule, please cite:
@inproceedings{ning2021multi,
title={Multi-Anchor Active Domain Adaptation for Semantic Segmentation},
author={Ning, Munan and Lu, Donghuan and Wei, Dong and Bian, Cheng and Yuan, Chenglang and Yu, Shuang and Ma, Kai and Zheng, Yefeng},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
pages={9112--9122},
year={2021}
}
Notes
The anchors are calcuated based on features captured by decoders.
In this paper, we utilize the more powerful decoder in DeeplabV3+, it may cause somewhere unfair. So we strongly recommend the ProDA which utilize origin DeeplabV2 decoder.