This repository is for PointASNL introduced in the following paper
Xu Yan, Chaoda Zheng, Zhen Li*, Sheng Wang and Shuguang Cui, "PointASNL: Robust Point Clouds Processing using Nonlocal Neural Networks with Adaptive Sampling", CVPR 2020 [arxiv].
If you find our work useful in your research, please consider citing:
@inproceedings{yan2020pointasnl,
title={Pointasnl: Robust point clouds processing using nonlocal neural networks with adaptive sampling},
author={Yan, Xu and Zheng, Chaoda and Li, Zhen and Wang, Sheng and Cui, Shuguang},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={5589--5598},
year={2020}
}
Clone the repository:
git clone https://github.com/yanx27/PointASNL.git
Installation instructions for Ubuntu 16.04 (available at CUDA10):
-
Make sure CUDA and cuDNN are installed. Only this configurations has been tested:
- Python 3.6.9, TensorFlow 1.13.1, CUDA 10.1
-
Compile the customized Tensorflow operators by
sh complile_op.sh
. N.B. If you installed Tensorflow in a virtual environment, it needs to be activated when running these scripts
Aligned ModelNet40 dataset can be found here. Since the randomness of data augmentation, the result of this code maybe slightly different from the result in paper, but it should be around 93%.
It will cost relatively long time in first epoch for cache construction.
# Training
$ python train.py --data [MODELNET40 PATH] --exp_dir PointASNL_without_noise
# Evaluation
$ python test.py --data [MODELNET40 PATH] --model_path log/PointASNL_without_noise/best_model.ckpt
Model with AS module is extremely robust for noisy data. You can use adaptive sampling by setting --AS
.
# Training
$ python train.py --data [MODELNET40 PATH] --exp_dir PointASNL_with_noise --AS
# Evaluation on noisy data
$ python test.py --data [MODELNET40 PATH] --model_path log/PointASNL_with_noise/best_model.ckpt --AS --noise
We provide two options for training on ScanNet dataset (with or without pre/post processing). With grid sampling processing, more input points and deeper network structure, our PointASNL can achieve 66.6% on ScanNet benchmark.
Official ScanNet dataset can be downloaded here.
If you choose training without grid sampling, you need firstly run ScanNet/prepare_scannet.py
, otherwise you can skip to training step.
This method converges relatively slower, and will achieve result around 63%.
# Training
$ cd ScanNet/
$ python train_scannet.py --data [SCANNET PATH] --log_dir PointASNL
# Evaluation
$ cd ScanNet/
$ python test_scannet.py --data [SCANNET PATH] --model_path log/PointASNL/latest_model.ckpt
We highly recommend training with this method, although it takes a long time to process the raw data, it can achieve results around 66% and will be faster to converge. Grid sampling pre-processing will be automatically conducted before training.
# Training
$ cd ScanNet/
$ python train_scannet_grid.py --data [SCANNET PATH] --log_dir PointASNL_grid --num_point 10240 --model pointasnl_sem_seg_res --in_radius 2
# Evaluation
$ cd ScanNet/
$ python test_scannet_grid.py --data [SCANNET PATH] --model_path log/PointASNL_grid/latest_model.ckpt
Model | mIoU | Download |
---|---|---|
pointasnl_sem_seg_res | 66.93 | ckpt-163.9M |
- SemanticKITTI dataset can be found here. Download the files related to semantic segmentation and extract everything into the same folder.
- We add codes with grid sampling processing, which can achieve better result of around 52% (using
--prepare_data
just in the first running). - Please using official semantic_kitti_api for evaluation.
# Training
$ cd SemanticKITTI/
$ python train_semantic_kitti.py --data [SemanticKITTI PATH] --log_dir PointASNL --with_remission
# or
$ python train_semantic_kitti_grid.py --data [SemanticKITTI PATH] --log_dir PointASNL_grid --prepare_data
# Evaluation
$ cd SemanticKITTI/
$ python test_semantic_kitti.py --data [SemanticKITTI PATH] --model_path log/PointASNL/latest_model.ckpt --with_remission
# or
$ python test_semantic_kitti_grid.py --data [SemanticKITTI PATH] --model_path log/PointASNL_grid/best_model.ckpt --test_area [e.g., 08]
- The original code is borrowed from PointNet++ and PointConv.
- The code with grid sampling is borrowed from KPConv and RandLA-Net.
- The kd-tree tool is from nanoflann.
This repository is released under MIT License (see LICENSE file for details).