DADA: Depth-aware Domain Adaptation in Semantic Segmentation

Updates

02/2020: Using CycleGAN translated images, The DADA model achieves (43.1%) on SYNTHIA-2-Cityscapes

Paper

DADA: Depth-aware Domain Adaptation in Semantic Segmentation
Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, Patrick Pérez
valeo.ai, France
IEEE International Conference on Computer Vision (ICCV), 2019

If you find this code useful for your research, please cite our paper:

@inproceedings{vu2019dada,
  title={DADA: Depth-aware Domain Adaptation in Semantic Segmentation},
  author={Vu, Tuan-Hung and Jain, Himalaya and Bucher, Maxime and Cord, Mathieu and P{\'e}rez, Patrick},
  booktitle={ICCV},
  year={2019}
}

Abstract

Unsupervised domain adaptation (UDA) is important for applications where large scale annotation of representative data is challenging. For semantic segmentation in particular, it helps deploy on real "target domain" data models that are trained on annotated images from a different "source domain", notably a virtual environment. To this end, most previous works consider semantic segmentation as the only mode of supervision for source domain data, while ignoring other, possibly available, information like depth. In this work, we aim at exploiting at best such a privileged information while training the UDA model. We propose a unified depth-aware UDA framework that leverages in several complementary ways the knowledge of dense depth in the source domain. As a result, the performance of the trained semantic segmentation model on the target domain is boosted. Our novel approach indeed achieves state-of-the-art performance on different challenging synthetic-2-real benchmarks.

Preparation

Pre-requisites

Python 3.7
Pytorch >= 1.2.0
CUDA 10.0 or higher
The latest version of the ADVENT code.

Installation

Install OpenCV if you don't already have it:

$ conda install -c menpo opencv

Install ADVENT, the latest version:

$ git clone https://github.com/valeoai/ADVENT.git
$ pip install -e ./ADVENT

Clone and install the repo:

$ git clone https://github.com/valeoai/DADA
$ pip install -e ./DADA

With this (the -e option of pip), you can edit the DADA code on the fly and import function and classes of DADA in other project as well.

Optional. To uninstall this package, run:

$ pip uninstall DADA

You can take a look at the Dockerfile if you are uncertain about steps to install this project.

Datasets

By default, the datasets are put in DADA/data. We use symlinks to hook the DADA codebase to the datasets. An alternative option is to explicitlly specify the parameters DATA_DIRECTORY_SOURCE and DATA_DIRECTORY_TARGET in YML configuration files.

SYNTHIA: Please first follow the instructions here to download the images. In this work, we used the SYNTHIA-RAND-CITYSCAPES (CVPR16) split. The segmentation labels can be found here. The dataset directory should have this basic structure:
```
DADA/data/SYNTHIA                           % SYNTHIA dataset root
├── RGB
├── parsed_LABELS
└── Depth
```
Cityscapes: Please follow the instructions in Cityscape to download the images and validation ground-truths. The Cityscapes dataset directory should have this basic structure:
```
DADA/data/Cityscapes                       % Cityscapes dataset root
├── leftImg8bit
│   ├── train
│   └── val
└── gtFine
    └── val
```
Mapillary: Please follow the instructions in Mapillary to download the images and validation ground-truths. The Mapillary dataset directory should have this basic structure:
```
DADA/data/mapillary                        % Mapillary dataset root
├── train
│   └── images
└── validation
    ├── images
    └── labels
```

Pre-trained models

Pre-trained models can be downloaded here and put in DADA/pretrained_models

Running the code

For evaluating pretrained networks, execute:

$ cd DADA/dada/scripts
$ python test.py --cfg ./<configs_dir>/dada_pretrained.yml
$ python test.py --cfg ./<configs_dir>/dada_cyclegan_pretrained.yml

<configs_dir> could be set as configs_s2c (SYNTHIA2Cityscapes) or configs_s2m (SYNTHIA2Mapillary)

Training

For the experiments done in the paper, we used pytorch 1.2.0 and CUDA 10.0. To ensure reproduction, the random seed has been fixed in the code. Still, you may need to train a few times or to train longer (by changing MAX_ITERS and EARLY_STOP) to reach the comparable performance.

By default, logs and snapshots are stored in DADA/experiments with this structure:

DADA/experiments
├── logs
└── snapshots

To train DADA:

$ cd DADA/dada/scripts
$ python train.py --cfg ./<configs_dir>/dada.yml
$ python train.py --cfg ./<configs_dir>/dada.yml --tensorboard         % using tensorboard

To train AdvEnt baseline:

$ cd DADA/dada/scripts
$ python train.py --cfg ./<configs_dir>/advent.yml
$ python train.py --cfg ./<configs_dir>/advent.yml --tensorboard         % using tensorboard

Testing

To test DADA:

$ cd DADA/dada/scripts
$ python test.py --cfg ./<configs_dir>/dada.yml

Acknowledgements

This codebase heavily depends on AdvEnt.

License

DADA is released under the Apache 2.0 license.

valeoai / DADA