Skip to content

ICRA 2019 "Self-supervised Sparse-to-Dense: Self-supervised Depth Completion from LiDAR and Monocular Camera"

License

Notifications You must be signed in to change notification settings

fangchangma/self-supervised-depth-completion

Repository files navigation

self-supervised-depth-completion

This repo is the PyTorch implementation of our ICRA'19 paper on "Self-supervised Sparse-to-Dense: Self-supervised Depth Completion from LiDAR and Monocular Camera", developed by Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman at MIT. A video demonstration is available on YouTube.

photo not available

Our network is trained with the KITTI dataset alone, without pretraining on Cityscapes or other similar driving dataset (either synthetic or real). The use of additional data is likely to further improve the accuracy.

Please create a new issue for code-related questions.

Contents

  1. Dependency
  2. Data
  3. Trained Models
  4. Commands
  5. Citation

Dependency

This code was tested with Python 3 and PyTorch 1.0 on Ubuntu 16.04.

pip install numpy matplotlib Pillow
pip install torch torchvision # pytorch

# for self-supervised training requires opencv, along with the contrib modules
pip install opencv-contrib-python==3.4.2.16

Data

  • Download the KITTI Depth Dataset from their website. Use the following scripts to extract corresponding RGB images from the raw dataset.
./download/rgb_train_downloader.sh
./download/rgb_val_downloader.sh

The downloaded rgb files will be stored in the ../data/data_rgb folder. The overall code, data, and results directory is structured as follows (updated on Oct 1, 2019)

.
├── self-supervised-depth-completion
├── data
|   ├── data_depth_annotated
|   |   ├── train
|   |   ├── val
|   ├── data_depth_velodyne
|   |   ├── train
|   |   ├── val
|   ├── depth_selection
|   |   ├── test_depth_completion_anonymous
|   |   ├── test_depth_prediction_anonymous
|   |   ├── val_selection_cropped
|   └── data_rgb
|   |   ├── train
|   |   ├── val
├── results

Trained Models

Download our trained models at http://datasets.lids.mit.edu/self-supervised-depth-completion to a folder of your choice.

Commands

A complete list of training options is available with

python main.py -h

For instance,

# train with the KITTI semi-dense annotations, rgbd input, and batch size 1
python main.py --train-mode dense -b 1 --input rgbd

# train with the self-supervised framework, not using ground truth
python main.py --train-mode sparse+photo 

# resume previous training
python main.py --resume [checkpoint-path] 

# test the trained model on the val_selection_cropped data
python main.py --evaluate [checkpoint-path] --val select

Citation

If you use our code or method in your work, please cite the following:

@article{ma2018self,
	title={Self-supervised Sparse-to-Dense: Self-supervised Depth Completion from LiDAR and Monocular Camera},
	author={Ma, Fangchang and Cavalheiro, Guilherme Venturelli and Karaman, Sertac},
	booktitle={ICRA},
	year={2019}
}
@article{Ma2017SparseToDense,
	title={Sparse-to-Dense: Depth Prediction from Sparse Depth Samples and a Single Image},
	author={Ma, Fangchang and Karaman, Sertac},
	booktitle={ICRA},
	year={2018}
}

Releases

No releases published

Packages

No packages published