Official implementation of "Depth Estimation from Monocular Images and Sparse Radar using Deep Ordinal Regression Network" (https://arxiv.org/abs/2107.07596), an accepted paper of ICIP2021.
Please check Dockerfile for environment settings and python packages
Use gen_interpolation.py to generate sparse depth, dense depth and height-extended radar depth for training and evaluation. Generated depths will be saved in the same data_root directory of nuScenes sample, and please remember to modify the dir path in the .py. The number of CPU cores is set to 25 for faster data generating.
pypardiso is highly recommended, otherwise one could have similar result via spsolve in scipy but with much slower speed. Interpolation code is from https://gist.github.com/ialhashim/be6235489a9c43c6d240e8331836586a
Please download the pretrained ResNet101 weight file from http://sceneparsing.csail.mit.edu/model/pretrained_resnet/resnet101-imagenet.pth and modify the weight path in model/resnet.py.
After modifying some paths in train_nusc.py and trian_nusc_radar.py, directly call the .py files for training the baseline model and the proposed model on nuScenes.
cd pretrained_weight and execute sh download_pretrained_weight.sh.
cd .. back to the root dir and execute python evaluate.py to validate the pretrained model on the val_list.
The train_scene and val_scene txt files in ./list/nusc/ are the train/val splits I used in the paper. The reason why I used my own splits instead of official train/val splits was just because I didn't realize there are official ones at the time I conducted the experiments.
If you find this work useful in your research, please consider citing:
@inproceedings{DORN_radar,
author={Lo, Chen-Chou and Vandewalle, Patrick},
booktitle={Proceedings of the IEEE International Conference on Image Processing},
title={Depth Estimation From Monocular Images And Sparse Radar Using Deep Ordinal Regression Network},
year={2021},
pages={3343-3347},
doi={10.1109/ICIP42928.2021.9506550}
}