[Project Page] Accepted in NeurIPS 2020 as Spotlight

Wasserstein Distances for Stereo Disparity Estimation

by Divyansh Garg, Yan Wang, Bharath Hariharan, Mark Campbell, Kilian Q. Weinberger and Wei-Lun Chao


  title={Wasserstein Distances for Stereo Disparity Estimation},
  author={Garg, Divyansh and Wang, Yan and Hariharan, Bharath and Campbell, Mark and Weinberger, Kilian and Chao, Wei-Lun},


Existing approaches to depth or disparity estimation output a distribution over a
set of pre-defined discrete values. This leads to inaccurate results when the true
depth or disparity does not match any of these values. The fact that this distribution
is usually learned indirectly through a regression loss causes further problems in
ambiguous regions around object boundaries. We address these issues using a new
neural network architecture that is capable of outputting arbitrary depth values,
and a new loss function that is derived from the Wasserstein distance between
the true and the predicted distributions. We validate our approach on a variety
of tasks, including stereo disparity and depth estimation, and the downstream 3D
object detection. Our approach drastically reduces the error in ambiguous regions,
especially around object boundaries that greatly affect the localization of objects in
3D, achieving the state-of-the-art in 3D object detection for autonomous driving.


  • 1 June 2021: Released all pretrained models. Added script to download KITTI Object Detection dataset.


Our Wasserstein loss W_loss can be easily plugged in existing stereo depth models to improve the training and obtain better results.

We release the code for CDN-PSMNet and CDN-SDN models.


  1. Python 3.7
  2. Pytorch 1.2.0+
  3. CUDA
  4. pip install -r ./requirements.txt
  5. SceneFlow
  6. KITTI

Pretrained Models

Place the checkpoint folders in ./results.

Depth Models

Disparity Models


You have to download the SceneFlow and KITTI datasets. The structures of the datasets are shown in below.

KITTI can be automatically downloaded using ./scripts/download_kitti.sh

SceneFlow Dataset Structure

    | monkaa
        | frames_cleanpass
        | disparity
    | driving
        | frames_cleanpass
        | disparity
    | flyingthings3d
        | frames_cleanpass 
        | disparity

KITTI Object Detection Dataset Structure

    | training
        | calib
        | image_2
        | image_3
        | velodyne
    | testing
        | calib
        | image_2
        | image_3

Generate soft-links of SceneFlow Datasets. The results will be saved in ./sceneflow folder. Please change to fakepath path-to-SceneFlow to the SceneFlow dataset location before running the script.

python sceneflow.py --path path-to-SceneFlow --force

Convert the KITTI velodyne ground truths to depth maps. Please change to fakepath path-to-KITTI to the SceneFlow dataset location before running the script.

python ./src/preprocess/generate_depth_map.py --data_path path-to-KITTI/ --split_file ./split/trainval.txt

Optionally download KITTI2015 datasets for evaluating stereo disparity models.

Training and Inference

We have provided all pretrained models Pretrained Models. If you only want to generate the predictions, you can directly go to step 3.

We use config files to simplify argument parsing. The default setting requires four gpus to train. You can use smaller batch sizes which are btrain and bval, if you don't have enough gpus.

We provide code for both stereo disparity and stereo depth models.

We optionally use Losswise to visualize training metrics. An API key can be obtained and added to a config key to enable it.

1 Train CDN-SDN from Scratch on SceneFlow Dataset

python ./src/main_depth.py -c src/configs/sceneflow_w1.config

The checkpoints are saved in ./results/stack_sceneflow_w1/.

Follow same procedure to train stereo disparity model, but use src/main_disp.py and change to a disparity config.

2 Train CDN-SDN on KITTI Dataset

python ./src/main_depth.py -c src/configs/kitti_w1.config \
    --pretrain ./results/sceneflow_w1/checkpoint.pth.tar --datapath  path-to-KITTI/training/

Before running, please change the fakepath path-to-KITTI/ to the correct one. --pretrain is the path to the pretrained model on SceneFlow. The training results are saved in ./results/kitti_w1_train.

If you are working on evaluating CDN on KITTI testing set, you might want to train CDN on training+validation sets. The training results will be saved in ./results/sdn_kitti_trainval.

python ./src/main_depth.py -c src/configs/kitti_w1.config \
    --pretrain ./results/sceneflow_w1/checkpoint.pth.tar \
    --datapath  path-to-KITTI/training/ --split_train ./split/trainval.txt \
    --save_path ./results/sdn_kitti_trainval

The disparity models can also be trained on KITTI2015 datasets using src/kitti2015_w1_disp.config.

3 Generate Predictions

Please change the fakepath path-to-KITTI. Moreover, if you use the our provided checkpoint, please modify the value of --resume to the checkpoint location.

  • a. Using the model trained on KITTI training set, and generating predictions on training + validation sets.
python ./src/main_depth.py -c src/configs/kitti_w1.config \
    --resume ./results/sdn_kitti_train/checkpoint.pth.tar --datapath  path-to-KITTI/training/ \
    --data_list ./split/trainval.txt --generate_depth_map --data_tag trainval

The results will be saved in ./results/sdn_kitti_train/depth_maps_trainval/.

  • b. Using the model trained on KITTI training + validation set, and generating predictions on testing sets. You will use them when you want to submit your results to the leaderboard.

The results will be saved in ./results/sdn_kitti_trainval_set/depth_maps_trainval/.

# testing sets
python ./src/main_depth.py -c src/configs/kitti_w1.config \
    --resume ./results/sdn_kitti_trainval/checkpoint.pth.tar --datapath  path-to-KITTI/testing/ \
    --data_list=./split/test.txt --generate_depth_map --data_tag test

The results will be saved in ./results/sdn_kitti_trainval/depth_maps_test/.

4 Train 3D Detection with Pseudo-LiDAR

For training 3D object detection models, follow step 4 and after in the Pseudo-LiDAR_V2 repo https://github.com/mileyan/Pseudo_Lidar_V2.


Results on the Stereo Disparity


3D Object Detection Results on KITTI leader board



Please feel free to email us if you have any questions.

Divyansh Garg ([email protected]),
Yan Wang ([email protected]),
Wei-Lun Chao ([email protected])