Domain adaptive training for semantic segmentation network using multiple source datasets.

• NVIDIA GPU with at least xx GB memory
• nvidia-docker

Dockerfile provided in this repo uses nvcr.io/nvidia/pytorch image as a base image.

make build # Built with nvcr.io/nvidia/pytorch:22.11-py3

Optionally, you can specify the version of base image.

make build VERSION=21.09-py3 # Built with nvcr.io/nvidia/pytorch:21.09-py3

Note: There could be some compatibility issues when using different versions than 22.11-py3. I prepared Dockerimage for building an image from 20.03-py3 with limited libraries (Dockerfile_2003). To build it, run the following command:

make build-2003

Train a model for each source dataset using an ordinary supervised learning.

make train-<source name> \ # source name: camvid, cityscapes, or forest
    VERSION=<version>

make generate-pseudo-labels VERSION=<version>

Pseudo-labels will be generated under pseudo_labels directory. To change parameters, modify scripts/generate_pseudo_labels.sh.

make train-greenhouse-pseudo-soft # Using soft pseudo-labels
make train-greenhouse-pseudo-hard # Using hard pseudo-labels

To change parameters, modify scripts/train_greenhouse_soft_pseudo.sh / scripts/train_greenhouse_hard_pseudo.sh.

Paper

@inproceedings{
  author  = {Matsuzaki, Shigemichi and Masuzawa, Hiroaki and Miura, Jun},
  arxivid = {2303.00979},
  title   = {{Multi-Source Soft Pseudo-Label Learning with Domain Similarity-based Weighting for Semantic Segmentation}},
  url     = {http://arxiv.org/abs/2303.00979},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year    = {2023}
}