Official PyTorch implementation of the CVPR 2020 paper "Autolabeling 3D Objects With Differentiable Rendering of SDF Shape Priors" by the ML Team at Toyota Research Institute (TRI), cf. References below. [Full paper] [YouTube]
![](/media/figs/tri-logo.png)
![](/media/figs/sdflabel-teaser.gif)
To set up the environment using conda, use the following commands:
conda env create -n sdflabel -f environment.yml
conda activate sdflabel
Add the sdfrenderer directory to PYTHONPATH:
export PYTHONPATH="${PYTHONPATH}:/path/to/sdfrenderer"
To run the optimization demo, first download the data folder. Then, extract the archive to the root folder of the project and run the following command:
python main.py configs/config_refine.ini --demo
To train the CSS network, run the following command:
python main.py configs/config_train.ini --train
The dataset of crops represents a collection of detected RGB patches (CSS input), corresponding NOCS patches (CSS output), and a JSON DB file comprising the patch relevant information (most importantly SDF latent vectors corresponding to the depicted 3D models). An example of such dataset is located in the data/db/crops folder.
Download KITTI 3D and modify the kitti_path in the config file config_refine.ini accordingly. To run optimization on the KITTI 3D dataset, run the following command:
python main.py configs/config_refine.ini --refine
Upon completion, autolabels will be stored to the output folder specified in the config file (output -> labels). To evaluate the generated dump, run:
python main.py configs/config_refine.ini --evaluate
The source code is released under the MIT license.
Sergey Zakharov*, Wadim Kehl*, Arjun Bhargava, Adrien Gaidon
@inproceedings{sdflabel,
author = {Sergey Zakharov and Wadim Kehl and Arjun Bhargava and Adrien Gaidon},
title = {Autolabeling 3D Objects with Differentiable Rendering of SDF Shape Priors},
booktitle = {IEEE Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}