## Mask3D: Mask Transformer for 3D Instance Segmentation
Jonas Schult1, Francis Engelmann2,3, Alexander Hermans1, Or Litany4, Siyu Tang3, Bastian Leibe1 1RWTH Aachen University 2ETH AI Center 3ETH Zurich 4NVIDIA Mask3D predicts accurate 3D semantic instances achieving state-of-the-art on ScanNet, ScanNet200, S3DIS and STPLS3D. [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/mask3d-for-3d-semantic-instance-segmentation/3d-instance-segmentation-on-scannetv2)](https://paperswithcode.com/sota/3d-instance-segmentation-on-scannetv2?p=mask3d-for-3d-semantic-instance-segmentation) [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/mask3d-for-3d-semantic-instance-segmentation/3d-instance-segmentation-on-scannet200)](https://paperswithcode.com/sota/3d-instance-segmentation-on-scannet200?p=mask3d-for-3d-semantic-instance-segmentation) [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/mask3d-for-3d-semantic-instance-segmentation/3d-instance-segmentation-on-s3dis)](https://paperswithcode.com/sota/3d-instance-segmentation-on-s3dis?p=mask3d-for-3d-semantic-instance-segmentation) [![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/mask3d-for-3d-semantic-instance-segmentation/3d-instance-segmentation-on-stpls3d)](https://paperswithcode.com/sota/3d-instance-segmentation-on-stpls3d?p=mask3d-for-3d-semantic-instance-segmentation) PyTorch Lightning Config: Hydra ![teaser](./docs/teaser.jpg)


[[Project Webpage](https://jonasschult.github.io/Mask3D/)] [[Paper](https://arxiv.org/abs/2210.03105)] [[Demo](https://francisengelmann.github.io/mask3d/)] ## News * **29. October 2023**: Check out this [easy setup](https://github.com/cvg/Mask3D) for Mask3D. * **17. January 2023**: Mask3D is accepted at ICRA 2023. :fire: * **14. October 2022**: STPLS3D support added. * **10. October 2022**: Mask3D ranks 2nd on the [STPLS3D Challenge](https://codalab.lisn.upsaclay.fr/competitions/4646#results) hosted by the [Urban3D Workshop](https://urban3dchallenge.github.io/) at ECCV 2022. * **6. October 2022**: [Mask3D preprint](https://arxiv.org/abs/2210.03105) released on arXiv. * **25. September 2022**: Code released. ## Code structure We adapt the codebase of [Mix3D](https://github.com/kumuji/mix3d) which provides a highly modularized framework for 3D Semantic Segmentation based on the MinkowskiEngine. ``` ├── mix3d │ ├── main_instance_segmentation.py <- the main file │ ├── conf <- hydra configuration files │ ├── datasets │ │ ├── preprocessing <- folder with preprocessing scripts │ │ ├── semseg.py <- indoor dataset │ │ └── utils.py │ ├── models <- Mask3D modules │ ├── trainer │ │ ├── __init__.py │ │ └── trainer.py <- train loop │ └── utils ├── data │ ├── processed <- folder for preprocessed datasets │ └── raw <- folder for raw datasets ├── scripts <- train scripts ├── docs ├── README.md └── saved <- folder that stores models and logs ``` ### Dependencies :memo: The main dependencies of the project are the following: ```yaml python: 3.10.9 cuda: 11.3 ``` You can set up a conda environment as follows ``` # Some users experienced issues on Ubuntu with an AMD CPU # Install libopenblas-dev (issue #115, thanks WindWing) # sudo apt-get install libopenblas-dev export TORCH_CUDA_ARCH_LIST="6.0 6.1 6.2 7.0 7.2 7.5 8.0 8.6" conda env create -f environment.yml conda activate mask3d_cuda113 pip3 install torch==1.12.1+cu113 torchvision==0.13.1+cu113 --extra-index-url https://download.pytorch.org/whl/cu113 pip3 install torch-scatter -f https://data.pyg.org/whl/torch-1.12.1+cu113.html pip3 install 'git+https://github.com/facebookresearch/detectron2.git@710e7795d0eeadf9def0e7ef957eea13532e34cf' --no-deps mkdir third_party cd third_party git clone --recursive "https://github.com/NVIDIA/MinkowskiEngine" cd MinkowskiEngine git checkout 02fc608bea4c0549b0a7b00ca1bf15dee4a0b228 python setup.py install --force_cuda --blas=openblas cd .. git clone https://github.com/ScanNet/ScanNet.git cd ScanNet/Segmentator git checkout 3e5726500896748521a6ceb81271b0f5b2c0e7d2 make cd ../../pointnet2 python setup.py install cd ../../ pip3 install pytorch-lightning==1.7.2 ``` ### Data preprocessing :hammer: After installing the dependencies, we preprocess the datasets. #### ScanNet / ScanNet200 First, we apply Felzenswalb and Huttenlocher's Graph Based Image Segmentation algorithm to the test scenes using the default parameters. Please refer to the [original repository](https://github.com/ScanNet/ScanNet/tree/master/Segmentator) for details. Put the resulting segmentations in `./data/raw/scannet_test_segments`. ``` python -m datasets.preprocessing.scannet_preprocessing preprocess \ --data_dir="PATH_TO_RAW_SCANNET_DATASET" \ --save_dir="data/processed/scannet" \ --git_repo="PATH_TO_SCANNET_GIT_REPO" \ --scannet200=false/true ``` #### S3DIS The S3DIS dataset contains some smalls bugs which we initially fixed manually. We will soon release a preprocessing script which directly preprocesses the original dataset. For the time being, please follow the instructions [here](https://github.com/JonasSchult/Mask3D/issues/8#issuecomment-1279535948) to fix the dataset manually. Afterwards, call the preprocessing script as follows: ``` python -m datasets.preprocessing.s3dis_preprocessing preprocess \ --data_dir="PATH_TO_Stanford3dDataset_v1.2" \ --save_dir="data/processed/s3dis" ``` #### STPLS3D ``` python -m datasets.preprocessing.stpls3d_preprocessing preprocess \ --data_dir="PATH_TO_STPLS3D" \ --save_dir="data/processed/stpls3d" ``` ### Training and testing :train2: Train Mask3D on the ScanNet dataset: ```bash python main_instance_segmentation.py ``` Please refer to the [config scripts](https://github.com/JonasSchult/Mask3D/tree/main/scripts) (for example [here](https://github.com/JonasSchult/Mask3D/blob/main/scripts/scannet/scannet_val.sh#L15)) for detailed instructions how to reproduce our results. In the simplest case the inference command looks as follows: ```bash python main_instance_segmentation.py \ general.checkpoint='PATH_TO_CHECKPOINT.ckpt' \ general.train_mode=false ``` ## Trained checkpoints :floppy_disk: We provide detailed scores and network configurations with trained checkpoints. ### [S3DIS](http://buildingparser.stanford.edu/dataset.html) (pretrained on ScanNet train+val) Following PointGroup, HAIS and SoftGroup, we finetune a model pretrained on ScanNet ([config](./scripts/scannet/scannet_pretrain_for_s3dis.sh) and [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/s3dis/scannet_pretrained/scannet_pretrained.ckpt)). | Dataset | AP | AP_50 | AP_25 | Config | Checkpoint :floppy_disk: | Scores :chart_with_upwards_trend: | Visualizations :telescope: |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:| | Area 1 | 69.3 | 81.9 | 87.7 | [config](scripts/s3dis/s3dis_pretrained.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/s3dis/scannet_pretrained/area1_scannet_pretrained.ckpt) | [scores](./docs/detailed_scores/s3dis/scannet_pretrained/s3dis_area1_scannet_pretrained.txt) | [visualizations](https://omnomnom.vision.rwth-aachen.de/data/mask3d/visualizations/s3dis/scannet_pretrained/area_1/) | Area 2 | 44.0 | 59.5 | 66.5 | [config](scripts/s3dis/s3dis_pretrained.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/s3dis/scannet_pretrained/area2_scannet_pretrained.ckpt) | [scores](./docs/detailed_scores/s3dis/scannet_pretrained/s3dis_area2_scannet_pretrained.txt) | [visualizations](https://omnomnom.vision.rwth-aachen.de/data/mask3d/visualizations/s3dis/scannet_pretrained/area_2/) | Area 3 | 73.4 | 83.2 | 88.2 | [config](scripts/s3dis/s3dis_pretrained.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/s3dis/scannet_pretrained/area3_scannet_pretrained.ckpt) | [scores](./docs/detailed_scores/s3dis/scannet_pretrained/s3dis_area3_scannet_pretrained.txt) | [visualizations](https://omnomnom.vision.rwth-aachen.de/data/mask3d/visualizations/s3dis/scannet_pretrained/area_3/) | Area 4 | 58.0 | 69.5 | 74.9 | [config](scripts/s3dis/s3dis_pretrained.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/s3dis/scannet_pretrained/area4_scannet_pretrained.ckpt) | [scores](./docs/detailed_scores/s3dis/scannet_pretrained/s3dis_area4_scannet_pretrained.txt) | [visualizations](https://omnomnom.vision.rwth-aachen.de/data/mask3d/visualizations/s3dis/scannet_pretrained/area_4/) | Area 5 | 57.8 | 71.9 | 77.2 | [config](scripts/s3dis/s3dis_pretrained.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/s3dis/scannet_pretrained/area5_scannet_pretrained.ckpt) | [scores](./docs/detailed_scores/s3dis/scannet_pretrained/s3dis_area5_scannet_pretrained.txt) | [visualizations](https://omnomnom.vision.rwth-aachen.de/data/mask3d/visualizations/s3dis/scannet_pretrained/area_5/) | Area 6 | 68.4 | 79.9 | 85.2 | [config](scripts/s3dis/s3dis_pretrained.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/s3dis/scannet_pretrained/area6_scannet_pretrained.ckpt) | [scores](./docs/detailed_scores/s3dis/scannet_pretrained/s3dis_area6_scannet_pretrained.txt) | [visualizations](https://omnomnom.vision.rwth-aachen.de/data/mask3d/visualizations/s3dis/scannet_pretrained/area_6/) ### [S3DIS](http://buildingparser.stanford.edu/dataset.html) (from scratch) | Dataset | AP | AP_50 | AP_25 | Config | Checkpoint :floppy_disk: | Scores :chart_with_upwards_trend: | Visualizations :telescope: |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:| | Area 1 | 74.1 | 85.1 | 89.6 | [config](scripts/s3dis/s3dis_from_scratch.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/s3dis/from_scratch/area1_from_scratch.ckpt) | [scores](./docs/detailed_scores/s3dis/from_scratch/s3dis_area1_from_scratch.txt) | [visualizations](https://omnomnom.vision.rwth-aachen.de/data/mask3d/visualizations/s3dis/from_scratch/area_1/) | Area 2 | 44.9 | 57.1 | 67.9 | [config](scripts/s3dis/s3dis_from_scratch.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/s3dis/from_scratch/area2_from_scratch.ckpt) | [scores](./docs/detailed_scores/s3dis/from_scratch/s3dis_area2_from_scratch.txt) | [visualizations](https://omnomnom.vision.rwth-aachen.de/data/mask3d/visualizations/s3dis/from_scratch/area_2/) | Area 3 | 74.4 | 84.4 | 88.1 | [config](scripts/s3dis/s3dis_from_scratch.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/s3dis/from_scratch/area3_from_scratch.ckpt) | [scores](./docs/detailed_scores/s3dis/from_scratch/s3dis_area3_from_scratch.txt) | [visualizations](https://omnomnom.vision.rwth-aachen.de/data/mask3d/visualizations/s3dis/from_scratch/area_3/) | Area 4 | 63.8 | 74.7 | 81.1 | [config](scripts/s3dis/s3dis_from_scratch.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/s3dis/from_scratch/area4_from_scratch.ckpt) | [scores](./docs/detailed_scores/s3dis/from_scratch/s3dis_area4_from_scratch.txt) | [visualizations](https://omnomnom.vision.rwth-aachen.de/data/mask3d/visualizations/s3dis/from_scratch/area_4/) | Area 5 | 56.6 | 68.4 | 75.2 | [config](scripts/s3dis/s3dis_from_scratch.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/s3dis/from_scratch/area5_from_scratch.ckpt) | [scores](./docs/detailed_scores/s3dis/from_scratch/s3dis_area5_from_scratch.txt) | [visualizations](https://omnomnom.vision.rwth-aachen.de/data/mask3d/visualizations/s3dis/from_scratch/area_5/) | Area 6 | 73.3 | 83.4 | 87.8 | [config](scripts/s3dis/s3dis_from_scratch.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/s3dis/from_scratch/area6_from_scratch.ckpt) | [scores](./docs/detailed_scores/s3dis/from_scratch/s3dis_area6_from_scratch.txt) | [visualizations](https://omnomnom.vision.rwth-aachen.de/data/mask3d/visualizations/s3dis/from_scratch/area_6/) ### [ScanNet v2](https://kaldir.vc.in.tum.de/scannet_benchmark/semantic_instance_3d?metric=ap) | Dataset | AP | AP_50 | AP_25 | Config | Checkpoint :floppy_disk: | Scores :chart_with_upwards_trend: | Visualizations :telescope: |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:| | ScanNet val | 55.2 | 73.7 | 83.5 | [config](scripts/scannet/scannet_val.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/scannet/scannet_val.ckpt) | [scores](./docs/detailed_scores/scannet_val.txt) | [visualizations](https://omnomnom.vision.rwth-aachen.de/data/mask3d/visualizations/scannet/val/) | ScanNet test | 56.6 | 78.0 | 87.0 | [config](scripts/scannet/scannet_benchmark.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/scannet/scannet_benchmark.ckpt) | [scores](http://kaldir.vc.in.tum.de/scannet_benchmark/result_details?id=1081) | [visualizations](https://omnomnom.vision.rwth-aachen.de/data/mask3d/visualizations/scannet/test/) ### [ScanNet 200](https://kaldir.vc.in.tum.de/scannet_benchmark/scannet200_semantic_instance_3d) | Dataset | AP | AP_50 | AP_25 | Config | Checkpoint :floppy_disk: | Scores :chart_with_upwards_trend: | Visualizations :telescope: |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:| | ScanNet200 val | 27.4 | 37.0 | 42.3 | [config](scripts/scannet200/scannet200_val.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/scannet200/scannet200_val.ckpt) | [scores](./docs/detailed_scores/scannet200_val.txt) | [visualizations](https://omnomnom.vision.rwth-aachen.de/data/mask3d/visualizations/scannet200/val/) | ScanNet200 test | 27.8 | 38.8 | 44.5 | [config](scripts/scannet200/scannet200_benchmark.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/scannet200/scannet200_benchmark.ckpt) | [scores](https://kaldir.vc.in.tum.de/scannet_benchmark/result_details?id=1242) | [visualizations](https://omnomnom.vision.rwth-aachen.de/data/mask3d/visualizations/scannet200/test/) ### [STPLS3D](https://www.stpls3d.com/) | Dataset | AP | AP_50 | AP_25 | Config | Checkpoint :floppy_disk: | Scores :chart_with_upwards_trend: | Visualizations :telescope: |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:| | STPLS3D val | 57.3 | 74.3 | 81.6 | [config](scripts/stpls3d/stpls3d_val.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/stpls3d/stpls3d_val.ckpt) | [scores](./docs/detailed_scores/stpls3d.txt) | [visualizations](https://omnomnom.vision.rwth-aachen.de/data/mask3d/visualizations/stpls3d/) | STPLS3D test | 63.4 | 79.2 | 85.6 | [config](scripts/stpls3d/stpls3d_benchmark.sh) | [checkpoint](https://omnomnom.vision.rwth-aachen.de/data/mask3d/checkpoints/stpls3d/stpls3d_benchmark.zip) | [scores](https://codalab.lisn.upsaclay.fr/competitions/4646#results) | visualizations ## BibTeX :pray: ``` @article{Schult23ICRA, title = {{Mask3D: Mask Transformer for 3D Semantic Instance Segmentation}}, author = {Schult, Jonas and Engelmann, Francis and Hermans, Alexander and Litany, Or and Tang, Siyu and Leibe, Bastian}, booktitle = {{International Conference on Robotics and Automation (ICRA)}}, year = {2023} } ```