Awesome
TriVol: Point Cloud Rendering via Triple Volumes
This is the source code of "TriVol: Point Cloud Rendering via Triple Volumes" (CVPR-2023).
<p float="left"> <img src="docs/trivol.png" width="700"/> </p>Requirements (Tested on 4 * RTX3090)
- Linux
- Python == 3.8
- Pytorch == 1.13.0
- CUDA == 11.7
Installation
Install from environment.yml
You can directly install the requirements through:
$ conda env create -f environment.yml
Or install packages seperately
-
Create Environment
$ conda create --name trivol python=3.8 $ conda activate trivol
-
Pytorch (Please first check your cuda version)
$ conda install pytorch==1.13.0 torchvision==0.14.0 torchaudio==0.13.0 pytorch-cuda=11.7 -c pytorch -c nvidia
-
$ pip install git+https://github.com/KAIR-BAIR/nerfacc.git
-
pip install torch-scatter -f https://data.pyg.org/whl/torch-1.13.0+cu117.html
-
Other python packages: open3d, opencv-python, etc.
Training
Training on the ShapeNet Car dataset
-
Dataset: Download our processed ShapeNet Car from HuggingFace.
-
Run: we recommand you run the code in multi-gpu mode.
$ conda activate trivol $ python train.py \ --scene_dir /path/to/shapenet \ --dataset shapenet \ --val_mode val \ --max_epochs 200 \ --lr 0.0005 \ --batch_size 1 \ --ngpus 4 \ --feat_dim 16 \ --patch_size 64 \ --exp_name shapenet_car
-
example results
<p float="left"> <img src="docs/samples_shapenet.jpg" width="600"/> </p>
Training on the ScanNet dataset
-
Dataset: Download and extract data from original ScanNet-V2 preprocess.
Dataset structure:
── scannet └── scene0000_00 ├── pose └──1.txt ├── intrinsic └──*.txt ├── color └──1.jpg └── scene0000_00_vh_clean_2.ply └── images.txt └── scene0000_01
-
Run: we recommand you run the code in multi-gpu mode.
$ conda activate trivol $ python train.py \ --scene_dir /path/to/scannet \ --dataset scannet \ --val_mode val \ --max_epochs 500 \ --lr 0.0005 \ --batch_size 1 \ --ngpus 4 \ --feat_dim 16 \ --patch_size 32 \ --exp_name scannet
-
example results
<p float="left"> <img src="docs/samples_scannet.jpg" width="600"/> </p>
Contact
Tao Hu - taohu@cse.cuhk.edu.hk
Citation
@InProceedings{Hu_2023_CVPR,
author = {Hu, Tao and Xu, Xiaogang and Chu, Ruihang and Jia, Jiaya},
title = {TriVol: Point Cloud Rendering via Triple Volumes},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2023},
pages = {20732-20741}
}