This is the PyTorch implementation for our BMVC20 (Oral) paper and our IJCV extension:

A. Lopez-Rodriguez, K. Mikolajczyk. DESC: Domain Adaptation for Depth Estimation via Semantic Consistency. [BMVC Paper] - [IJCV Extension]

Tested with Pytorch 1.4/1.5, CUDA 10.1, Ubuntu 18.04 and Python 3.6.9.

You need to install the Detectron2 library (used for semantic information) following these instructions. The pretrained panoptic segmentation model can be downloaded from here.

You also need to install OpenCV, ImageIO and SciPY, which can be done using:

pip install -r requirements.txt

KITTI

We provide the KITTI ground-truth depth maps for the eigen test split here in the file gt_depths.npz, which are generated using the export_gt_depth.py in the Monodepth2 repository.

vKITTI

We first train the networks separately by running the following two scripts

./pretrain_depth.sh VKITTI_ROOT_FOLDER KITTI_ROOT_FOLDER

./pretrain_semantic_depth.sh VKITTI_ROOT_FOLDER KITTI_ROOT_FOLDER

We then train them jointly to get our final model by using

./joint_training.sh VKITTI_ROOT_FOLDER KITTI_ROOT_FOLDER

Pretrained models for the depth estimation network can be found in this link. You need to have the ground-truth for the test data in the root folder, which is also given in the same link in gt_depths.npz as mentioned in the Datasets section.

To test the models we can run the following command

./test.sh KITTI_ROOT_FOLDER

By default it will load the model generated after finishing training, i.e, after running ./joint_training.sh. You can modify test.py to load the pretrained models, we give examples to do so in the commented lines. Also, if you are evaluating the stereo-trained model, set the disable_median_scaling option in evaluate_model to 1.

If you use DESC for your research, you can cite the paper using the following Bibtex entries:

@inproceedings{lopez2020desc,
  title={DESC: Domain Adaptation for Depth Estimation via Semantic Consistency},
  author={Lopez-Rodriguez, Adrian and Mikolajczyk, Krystian},
  booktitle={British Machine Vision Conference (BMVC)},
  year={2020}
}
@article{lopez2022desc,
  title={Desc: Domain adaptation for depth estimation via semantic consistency},
  author={Lopez-Rodriguez, Adrian and Mikolajczyk, Krystian},
  journal={International Journal of Computer Vision},
  pages={1--20},
  year={2022},
  publisher={Springer}
}

Our reported results for GASDA are better than those in the original GASDA paper due to an indexing bug the original GASDA code. The indexing bug was related to the test ground-truth generation from the Velodyne data, which has already been fixed in GASDA and now their results match those reported in our paper.

Code is inspired by T^2Net and GASDA.

Adrian Lopez-Rodriguez: [email protected]