Xiang Zhang*, Zeyuan Chen*, Fangyin Wei, and Zhuowen Tu (*Equal contribution)

This is the repository for the paper Uni-3D: A Universal Model for Panoptic 3D Scene Reconstruction (ICCV 2023).

[Paper]

We have pre-packaged all dependencies via Docker image. It is built on top of PyTorch 2.1.2 with CUDA 11.8.

You can pull the image via

docker pull zx1239856/uni-3d:0.1.0

Assume you already have proper PyTorch (>=1.10.1) and CUDA (>=11.3) installation.

1. Install the following system dependencies

apt-get install ninja-build libopenblas-dev libopenexr-dev

2. Remove the comment mark on Line 9 of requirements.txt. Install the required Python packages via

pip install -r requirements.txt

Please download 3D-FRONT from Dahnert et al. (Panoptic 3D Scene Reconstruction from a Single RGB Image). Extract it under datasets/front3d/data as

unzip front3d.zip -d datasets/front3d/data

Please request the dataset from the authors of Pano-Re. Extract it under datasets/matterport/data.

Also download the room mask and depth from BUOL. Extract them underdataset/matterport/room_mask and dataset/matterport/depth_gen, respectively.

matterport/
    meta/
        train_3d.json                                         # Training set metadata
        ...
    data/
        <scene_id>/            
            ├── <image_id>_i<frame_id>.png                    # Color image: 320x240x3
            ├── <image_id>_segmap<frame_id>.mapped.npz        # 2D Segmentation: 320x240x2, with 0: pre-mapped semantics, 1: instances
            ├── <image_id>_intrinsics_<camera_id>.png         # Intrinsics matrix: 4x4
            ├── <image_id>_geometry<frame_id>.npz             # 3D Geometry: 256x256x256x1, truncated, (unsigned) distance field at 3cm voxel resolution and 12 voxel truncation.
            ├── <image_id>_segmentation<frame_id>.mapped.npz  # 3D Segmentation: 256x256x256x2, with 0: pre-mapped semantics & instances
            ├── <image_id>_weighting<frame_id>.npz            # 3D Weighting mask: 256x256x256x1
    depth_gen/
        <scene_id>/     
            ├── <posithion_id>_d<frame_id>.png                # Depth image: 320x240x1
    room_mask/
        <scene_id>/   
            ├── <posithion_id>_rm<frame_id>.png               # Room mask: 320x240x1

If you are using docker, you may set the following prefix for convenience.

export DOCKER_PREFIX="docker run -it --gpus all --shm-size 128G -v "$(pwd)":/workspace zx1239856/uni-3d:0.1.0"

$DOCKER_PREFIX OMP_NUM_THREADS=16 torchrun --nproc_per_node=8 train_net.py --config-file configs/front3d/mask2former_R50_bs16_160k.yaml OUTPUT_DIR <path-to-output-dir>

$DOCKER_PREFIX OMP_NUM_THREADS=16 torchrun --nproc_per_node=8 train_net.py --config-file configs/front3d/uni_3d_R50.yaml MODEL.WEIGHTS <path-to-pretrained-2d-model> OUTPUT_DIR <path-to-output-dir>

Use uni_3d_R50_ms.yaml for multi-scale feature reprojection.

Please adjust --nproc_per_node, OMP_NUM_THREADS and SOLVER.IMS_PER_BATCH based on your environment.

Please add --eval-only flag to the training scripts above for evaluation.

You can generate meshes for visualization for 3D-FRONT images via the following command.

python demo_front3d.py -i <path-to-3d-front-image> -o <path-to-output-dir> -m <path-to-pretrained-model>

Please consider citing Uni-3D if you find the work helpful.

@InProceedings{Zhang_2023_ICCV,
    author    = {Zhang, Xiang and Chen, Zeyuan and Wei, Fangyin and Tu, Zhuowen},
    title     = {Uni-3D: A Universal Model for Panoptic 3D Scene Reconstruction},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {9256-9266}
}

This repository is released under the Apache License 2.0. License can be found in LICENSE file.

• Mask2Former for the framework.
• panoptic-reconstruction for the pre-processed 3D-FRONT and Matterport dataset, and evaluation codes.
• BUOL for generated depth and room mask on Matterport dataset.