Frugal_PN_Training

This is an implementation of our paper "Frugal 3D Point Cloud Model Training via Progressive Near Point Filtering and Fused Aggregation".

Install

source install.sh

Note:

We recommend using CUDA 11.x; check your CUDA version by: nvcc --version before using the bash file;

Usage

L-FPS

Can be easily applied to other models and datasets by changing few lines of dataloader. Please refer to openpoints/dataset/s3dis/s3dis.py, openpoints/dataset/scannetv2/scannet.py, and openpoints/dataset/modelnet/modelnet40_ply_2048_loader.py. We also register offlinefps operation in OpenPoints library. Implementation can be found in subsample.py and openpoints/cpp/pointnet2_batch/src/sampling_gpu.cu.

Fused Aggregation

We register fused_group_and_reduce and fused_group_and_reduce_pe operation in OpenPoints library. Implementation can be found in group.py and openpoints/cpp/pointnet2_batch/src. They can be used just like other operations in OpenPoints.

# Without positional encoding
idx = ball_query(self.radius, self.nsample, support_xyz, query_xyz)
features = fused_group_and_reduce(features, idx)

# With positional encoding
idx = ball_query(self.radius, self.nsample, support_xyz, query_xyz)
features = fused_group_and_reduce_pe(features, idx, grouped_pos_enc)

Prepare Dataset

ModelNet40

ModelNet40 dataset is downloaded automatically.

S3DIS

mkdir -p data/S3DIS/
cd data/S3DIS
gdown https://drive.google.com/uc?id=1MX3ZCnwqyRztG1vFRiHkKTz68ZJeHS4Y
tar -xvf s3disfull.tar

ScanNet

cd data
gdown https://drive.google.com/uc?id=1uWlRPLXocqVbJxPvA2vcdQINaZzXf1z_
tar -xvf ScanNet.tar

Extract L-FPS results

ModelNet40

CUDA_VISIBLE_DEVICES=0 python examples/classification/main.py --cfg cfgs/modelnet40ply2048/pointnet++da-lfps.yaml mode=extract_fps wandb.use_wandb=False

S3DIS

CUDA_VISIBLE_DEVICES=0 bash script/extract_fps_segmentation.sh cfgs/s3dis/pointmetabase-l-lfps.yaml wandb.use_wandb=False

ScanNet

CUDA_VISIBLE_DEVICES=0 bash script/extract_fps_segmentation.sh cfgs/scannet/pointmetabase-l-lfps.yaml wandb.use_wandb=False

Train

ModelNet40 (PointNet++)

# Run baseline
CUDA_VISIBLE_DEVICES=0 python examples/classification/main.py --cfg cfgs/modelnet40ply2048/pointnet++da.yaml wandb.use_wandb=False

# Run with L-FPS
CUDA_VISIBLE_DEVICES=0 python examples/classification/main.py --cfg cfgs/modelnet40ply2048/pointnet++da-lfps.yaml wandb.use_wandb=False

# Run with fused aggregation
CUDA_VISIBLE_DEVICES=0 python examples/classification/main.py --cfg cfgs/modelnet40ply2048/pointnet++da-fused.yaml wandb.use_wandb=False

# Run with all optimizations
CUDA_VISIBLE_DEVICES=0 python examples/classification/main.py --cfg cfgs/modelnet40ply2048/pointnet++da-all.yaml wandb.use_wandb=False

S3DIS, ScanNet (PointNet++, PointMetaBase)

# Run baseline
CUDA_VISIBLE_DEVICES=0 bash script/main_segmentation.sh cfgs/[s3dis, scannet]/[pointnet++da, pointmetabase-l, pointmetabase-xl].yaml wandb.use_wandb=False

# Run with L-FPS
CUDA_VISIBLE_DEVICES=0 bash script/main_segmentation.sh cfgs/[s3dis, scannet]/[pointnet++da, pointmetabase-l, pointmetabase-xl]-lfps.yaml wandb.use_wandb=False

# Run with fused aggregation
CUDA_VISIBLE_DEVICES=0 bash script/main_segmentation.sh cfgs/[s3dis, scannet]/[pointnet++da, pointmetabase-l, pointmetabase-xl]-fused.yaml wandb.use_wandb=False

# Run with all optimizations
CUDA_VISIBLE_DEVICES=0 bash script/main_segmentation.sh cfgs/[s3dis, scannet]/[pointnet++da, pointmetabase-l, pointmetabase-xl]-all.yaml wandb.use_wandb=False

Acknowledgment

This repository is built on reusing codes of PointMetaBase, OpenPoints and PointNeXt.

Citation

@inproceedings {frugaltraining,
    title={Frugal 3D Point Cloud Model Training via Progressive Near Point Filtering and Fused Aggregation},
    author={Donghyun Lee and Yejin Lee and Jae W. Lee and Hongil Yoon},
    booktitle = {European Conference on Computer Vision ({ECCV} 24)},
    year={2024},
}