Normal Inference Module

This is a PyTorch demo of the Normal Inference Module (NIM), presented in our IROS 2020 paper, Applying Surface Normal Information in Drivable Area and Road Anomaly Detection for Ground Mobile Robots. Our NIM can be used effectively for estimating surface normal information from depth images. The code has been tested in Python 3.6 and PyTorch 1.7.

We provide two examples in examples, where rgb, depth_u16 and calib contain RGB images, depth images and calibration files, respectively. These examples belong to the KITTI road dataset.

Run demo.py, and then the surface normal estimation will be saved in examples/normal. Please note that our NIM can run in two different ways. Set sign_filter=True, and then our NIM will additionally utilize a sign filter.

If you use this code for your research, please cite our paper.

@inproceedings{wang2020applying,
  title        = {Applying surface normal information in drivable area and road anomaly detection for ground mobile robots},
  author       = {Wang, Hengli and Fan, Rui and Sun, Yuxiang and Liu, Ming},
  booktitle    = {2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages        = {2706--2711},
  year         = {2020},
  organization = {IEEE},
  doi          = {10.1109/IROS45743.2020.9341340}
}