aim-uofa / DyCo3D

DyCo3d

DyCo3d: Robust Instance Segmentation of 3D Point Clouds through Dynamic Convolution (CVPR 2021)

Code for the paper DyCo3D: Robust Instance Segmentation of 3D Point Clouds through Dynamic Convolution, CVPR 2021.

Authors: Tong He, Chunhua Shen, Anton van den Hengel

[arxiv]

Introduction

Previous top-performing approaches for point cloud instance segmentation involve a bottom-up strategy, which often includes inefficient operations or complex pipelines, such as grouping over-segmented components, introducing additional steps for refining, or designing complicated loss functions. The inevitable variation in the instance scales can lead bottom-up methods to become particularly sensitive to hyper-parameter values. To this end, we propose instead a dynamic, proposal-free, data-driven approach that generates the appropriate convolution kernels to apply in response to the nature of the instances. To make the kernels discriminative, we explore a large context by gathering homogeneous points that share identical semantic categories and have close votes for the geometric centroids. Instances are then decoded by several simple convolutional layers. Due to the limited receptive field introduced by the sparse convolution, a small light-weight transformer is also devised to capture the long-range dependencies and high-level interactions among point samples. The proposed method achieves promising results on both ScanetNetV2 and S3DIS, and this performance is robust to the particular hyper-parameter values chosen. It also improves inference speed by more than 25% over the current state-of-the-art.

Installation

Requirements

• Python 3.7.0
• Pytorch 1.1.0
• CUDA 10.1

Virtual Environment

conda create -n dyco3d python==3.7
conda activate dyco3d

Install DyCo3d (Follow the installation steps of PointGroup)

(1) Clone the DyCo3d repository.

git clone https://github.com/aim-uofa/DyCo3D.git
cd DyCo3D

(2) Install the dependent libraries.

pip install -r requirements.txt
conda install -c bioconda google-sparsehash 

(3) For the SparseConv, we use the repo from PointGroup

• To compile spconv, firstly install the dependent libraries.

conda install libboost
conda install -c daleydeng gcc-5 # need gcc-5.4 for sparseconv

Add the $INCLUDE_PATH$ that contains boost in lib/spconv/CMakeLists.txt. (Not necessary if it could be found.)

include_directories($INCLUDE_PATH$)

• Compile the spconv library.

cd lib/spconv
python setup.py bdist_wheel

• Run cd dist and pip install the generated .whl file.

(4) Compile the pointgroup_ops library.

cd lib/pointgroup_ops
python setup.py develop

If any header files could not be found, run the following commands.

python setup.py build_ext --include-dirs=$INCLUDE_PATH$
python setup.py develop

$INCLUDE_PATH$ is the path to the folder containing the header files that could not be found.

Data Preparation

(1) Download the ScanNet v2 dataset.

(2) Put the data in the corresponding folders.

• Copy the files [scene_id]_vh_clean_2.ply, [scene_id]_vh_clean_2.labels.ply, [scene_id]_vh_clean_2.0.010000.segs.json and [scene_id].aggregation.json into the dataset/scannetv2/train and dataset/scannetv2/val folders according to the ScanNet v2 train/val split.

• Copy the files [scene_id]_vh_clean_2.ply into the dataset/scannetv2/test folder according to the ScanNet v2 test split.

• Put the file scannetv2-labels.combined.tsv in the dataset/scannetv2 folder.

The dataset files are organized as follows.

DyCo3D
├── dataset
│   ├── scannetv2
│   │   ├── train
│   │   │   ├── [scene_id]_vh_clean_2.ply & [scene_id]_vh_clean_2.labels.ply & [scene_id]_vh_clean_2.0.010000.segs.json & [scene_id].aggregation.json
│   │   ├── val
│   │   │   ├── [scene_id]_vh_clean_2.ply & [scene_id]_vh_clean_2.labels.ply & [scene_id]_vh_clean_2.0.010000.segs.json & [scene_id].aggregation.json
│   │   ├── test
│   │   │   ├── [scene_id]_vh_clean_2.ply 
│   │   ├── scannetv2-labels.combined.tsv

(3) Generate input files [scene_id]_inst_nostuff.pth for instance segmentation.

cd dataset/scannetv2
python prepare_data_inst.py --data_split train
python prepare_data_inst.py --data_split val
python prepare_data_inst.py --data_split test

You can also download the data here (about 13G). The dataset files are organized as

DyCo3D
├── dataset
│   ├── scannetv2
│   │   ├── train
│   │   ├── val
│   │   ├── test
│   │   ├── val_gt

Training

CUDA_VISIBLE_DEVICES=0,1,2 python -m torch.distributed.launch --nproc_per_node=3 --master_port=$((RANDOM + 10000)) train.py --config config/dyco3d_multigpu_scannet.yaml  --output_path OUTPUT_DIR  --use_backbone_transformer

Inference and Evaluation

To test with a pretrained model, run

CUDA_VISIBLE_DEVICES=0 python test.py --config config/dyco3d_multigpu_scannet.yaml --output_path exp/model --resume MODEL --use_backbone_transforme

Pretrained Model

We provide a pretrained model trained on ScanNet v2 dataset. Download it here. Its performance on ScanNet v2 validation set is 35.5/57.6/72.9 in terms of mAP/mAP50/mAP25. (with a masking head size of 16)

Results on ScanNet Benchmark

Quantitative results on ScanNet test set at the submisison time.

Citation

If you find this work useful in your research, please cite:

@inproceedings{He2021dyco3d,
  title     =   {{DyCo3d}: Robust Instance Segmentation of 3D Point Clouds through Dynamic Convolution},
  author    =   {Tong He and Chunhua Shen and Anton van den Hengel},
  booktitle =   {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      =   {2021}
}

Acknowledgement

This repo is built upon PointGroup, spconv, condinst.

Contact

If you have any questions or suggestions about this repo, please feel free to contact me ([e-mail]).