azuki-miho / occuseg

This is the official code repository for OccuSeg, a state-of-the-art method for accurate joint 3D semantic and instance segmentation.

OccuSeg: Occupancy-aware 3D Instance Segmentation

Introduction

This is the official code repository for OccuSeg, a state-of-the-art method for accurate joint 3D semantic and instance segmentation.

This is a project from LuVision SIGMA, Tsinghua University. Visit our website for more interesting works: http://www.luvision.net/

License

This project is released under the GPLv3 license. We only allow free use for academic use. For commercial use, please contact us to negotiate a different license by: fanglu at tsinghua.edu.cn

Citing

If you find our code useful, please kindly cite our paper:

@INPROCEEDINGS{9157103,
  author={Han, Lei and Zheng, Tian and Xu, Lan and Fang, Lu},
  booktitle={2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, 
  title={OccuSeg: Occupancy-Aware 3D Instance Segmentation}, 
  year={2020},
  volume={},
  number={},
  pages={2937-2946},
  doi={10.1109/CVPR42600.2020.00301}}
}

Quickstart with docker

0. Install docker and nvidia runtime following here.
1. Download the preprocessed ScanNet dataset from http://153.35.185.228:81/opensource_data/OccuSeg/scannet_data_occuseg.tar.gz
2. Modify the correct paths to the dataset and output directory in docker_run.sh
3. Train the model with bash docker_run.sh. Replace train_instance.sh with evaluate_instance.sh to perform evaluation. A pertrained model is available at http://153.35.185.228:81/opensource_data/OccuSeg/baseline_m64.pth, which is trained on the ScanNet training set.

If you prefer to setup the environment or prepare the data manually, following the below instructions or checkout the docker_build.sh.

Environment setup

Preliminary Requirements:

• Ubuntu 16.04
• CUDA 9.0

We trained using NVidia GV100.

Conda environment

Create the conda environment using:

conda env create -f p1.yml

and activate it.

Install dependencies and SparseConvNet

sh all_build.sh

Data preparation

1. ScanNet data split

Put the scannet data folder at examples/ScanNet/datasets/scannet_data/, organized as

scannet_data
|---- train
        |--- scenexxxx_xx
|---- val
        |--- scenexxxx_xx
|---- test
        |--- scenexxxx_xx

2. Perform super-voxel clustering

Use the Segmentator at:

https://github.com/ScanNet/ScanNet/tree/master/Segmentator

For each scenexxxx_xx_vh_clean_2.ply, we generate a scenexxxx_xx_vh_clean_2.regions.json using the Segmentator with default parameters.

3. Generate pth files

Under examples/ScanNet/, run

python prepare_data.py

This will generate pths at scannet_data/instance/. Then divide into train/val/test.

We use a full_train data split to train the final model for benchmark, which is described in examples/ScanNet/datasets/full_train.txt and examples/ScanNet/datasets/full_val.txt, showing the files in folder scannet_data/instance/full_train and scannet_data/instance/full_val respectively.

After this step, the scannet_data folder should look like this:

scannet_data
|---- instance
        |---- train
                |--- scenexxxx_xx_vh_clean_2_instance.pth
        |---- val
                |--- scenexxxx_xx_vh_clean_2_instance.pth
        |---- test
                |--- scenexxxx_xx_vh_clean_2_instance.pth
        |---- full_train
                |--- scenexxxx_xx_vh_clean_2_instance.pth
        |---- full_val
                |--- scenexxxx_xx_vh_clean_2_instance.pth

Training and testing

To train the model, go to ScanNet folder:

cd example/ScanNet
sh training_script/train_instance.sh

Training checkpoints are saved at ckpts.

To evaluate the results on test set:

sh training_script/evaluate_instance.sh

Make sure to modify the evaluate_instance.sh to select a checkpoint.

To evaluate on the val set, run:

sh training_script/evaluate_instance.sh