Pandora computes a disparity map from stereo rectified images
Pandora aims at shortening the path between a stereo-matching prototype and its industrialized version. By providing a modular pipeline inspired from the [Scharstein2002] taxonomy, it allows one to emulate, analyse and hopefully improve state of the art stereo algorithms with a few lines of code.
We (CNES) have actually been using Pandora to create the stereo matching pipeline for the CNES & Airbus CO3D off board processing chain. Leaning on Pandora's versatility and a fast-paced constantly evolving field we are still calling this framework a work in progress !
Pandora is available on Pypi and can be installed by:
pip install pandora #for the latest official release
For stereo reconstruction we invite you to install pandora and the required plugins using instead the following shortcut:
pip install pandora[sgm]
pip install pandora[mccnn]
Pandora requires a config.json to declare the pipeline and the stereo pair of images to process. Use our data_sample.zip to start right away !
pip install pandora #install pandora latest release
wget https://raw.githubusercontent.com/CNES/Pandora/master/data_samples/images/cones.zip # input stereo pair
wget https://raw.githubusercontent.com/CNES/Pandora/master/data_samples/json_conf_files/a_local_block_matching.json # configuration file
unzip cones.zip #uncompress data
pandora a_local_block_matching.json output_dir #run pandora
To create you own stereo matching pipeline and choose among the variety of algorithms we provide, please consult userguide
You will learn:
- which stereo matching steps you can use and combine
- how to quickly set up a Pandora pipeline
- how to add your own private algorithms to customize your Pandora Framework
- how to use Pandora API (see CARS for real life example)
Pandora uses transitions to manage the pipelines one can create. Images I/O are provided by rasterio and we use xarray to handle 3D Cost Volumes with few numba optimisations.
Our data test sample is based on the 2003 Middleburry dataset [Scharstein2003].
- Plugin_LibSGM - Stereo Matching Algorithm plugin for Pandora
- Plugin_MC-CNN - MC-CNN Neural Network plugin for Pandora
- CARS - CNES 3D reconstruction software
Please cite the following paper when using Pandora:
Cournet, M., Sarrazin, E., Dumas, L., Michel, J., Guinet, J., Youssefi, D., Defonte, V., Fardet, Q., 2020. Ground-truth generation and disparity estimation for optical satellite imagery. ISPRS - International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences.
- Scharstein2002
Scharstein D. & Szeliski R., 2002). Scharstein, D., & Szeliski, R. (2002). A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. International journal of computer vision, 47(1-3), 7-42.
- Scharstein2003
Scharstein D. & Szeliski R., 2003). Scharstein, D., & Szeliski, R. (2003, June). High-accuracy stereo depth maps using structured light. In 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings. (Vol. 1, pp. I-I). IEEE.