It contains 3196 object-scene configurations which involve 1966 objects and 500 indoor scenes. For each configurations, we generate following multi-view captures and corresponding G.T. object models.
- Images
- Depths
- Object masks
- Camera Poses
- 3D Model (pointcloud, SDF, water-tight mesh)
conda create --name fvor python=3.7
conda activate fvor
- Install blender 2.93.1
mkdir -p third_party && cd third_party && wget https://mirror.clarkson.edu/blender/release/Blender2.93/blender-2.93.1-linux-x64.tar.xz && tar -xvf blender-2.93.1-linux-x64.tar.xz && rm blender-2.93.1-linux-x64.tar.xz && cd ..
- Install habitat-sim. Clone habitat-sim repo using following commands and install it following their instructions.
conda install -y habitat-sim withbullet headless -c conda-forge -c aihabitat
mkdir -p third_party && cd third_party && git clone https://github.com/facebookresearch/habitat-sim.git && cd ../
- Install other package dependency
- (Optional) Setup TSDF fusion tools. Only necessary if you want to process the ABO assets yourself.
pip install cython
pip install open3d-python
pip install trimesh
pip install opencv-python==4.5.3.56
conda install -y -c conda-forge igl
cd libfusiongpu
mkdir build
cd build
cmake ..
make
cd ..
python setup.py build_ext --inplace
cd ..
cd libmcubes
python setup.py build_ext --inplace
cd ..
- Amazon-Berkeley Object Dataset Download abo-3dmodels.tar, put the abo-3dmodels.tar file in the data folder. Then uncompress with command
tar -xvf abo-3dmodels.tar
- Habitat - Matterport 3D Research Dataset Download the hm3d-train-glb.tar and hm3d-train-habitat.tar, put the hm3d-train-habitat.tar file and hm3d-train-glb.tar in the data folder. Then uncompress with commands
mkdir hm3d-train-glb && tar -xvf hm3d-train-glb.tar -C hm3d-train-glb
mkdir hm3d-train-habitat && tar -xvf hm3d-train-habitat.tar -C hm3d-train-habitat
- Our Provided Configurations Includes object placement and camera pose configurations.
- Sample Data contains one data object-scene configuration.
You should have a structure like these:
./data/
├── 3dmodels
│ ├── metadata
│ └── original
├── camera_pose_configs
│ └── ZxkSUELrWtQ_B07PC15YLQ.npy
├── hm3d-train-glb
│ └── 00459-Ze6tkhg7Wvc
└── hm3d-train-habitat
└── 00459-Ze6tkhg7Wvc
We provide configurations for object placement as well as camera poses. After download our configuration files, run following command to render images using Blender.
bash render.sh
Tunable parameters:
- The lens argument in render.sh could be adjusted to produce different FoV(field of view). The default lens(30) produce 62° horizontal FoV.
- The n_jobs_per_gpu could be lowered to accomodate your GPU memory.
Finally, your dataset structure should be looks like following:
data
└── HM3D_ABO
├── abo_assets
│ ├── B075X4YDCM.point.npz
│ └── B075X4YDCM.sdf.npz
└── scenes
└── 1S7LAXRdDqK_B075X4YDCM
├── rgb
├── pose
├── mask
├── depth
├── intrinsic.txt
└── obj_pose.txt
We've included our code for generating object and camera configurations. The basic idea is simple: we find a walkable locations of the scene, and try to place an object into that location and sample some camera around it. This section instruct you how to run our pipeline. Before we start, make sure that your have downloaded Amazon-Berkeley Object Dataset and Habitat - Matterport 3D Research Dataset. Make sure you've correctly setup the HM3D and ABO assets as instructed above.
Then, run following command to sample configurations.
bash sample.sh
The results are a set of {scene_id}_{model_id}.npy files inside camera_pose_configs folder.
Tunable parameters:
- configure camera_mode argument in sample.py for different type of camera poses.
- configure OUTPUT_DIR argument in sample.py.
Please first follow the installation instructions for Setup TSDF fusion tools. We generate two files for each ABO models. model_id.point.npz contains the surface point cloud and normals. model_id.sdf.npz contains sampled signed distance function.
To generate the signed distance function, we follow a similar procedure as Occupancy Network. We first render the depth map from a dense array of cameras surrounding the object. Then we use TSDF fusion to get the watertight mesh, which was used to calculate the signed distance function for sampled points.
bash process_abo.sh
If you use the our dataset in your research, please cite the following papers:
@article{yang2022hm3d,
title={HM3D-ABO: A Photo-realistic Dataset for Object-centric Multi-view 3D Reconstruction},
author={Yang, Zhenpei and Zhang, Zaiwei and Huang, Qixing},
journal={arXiv preprint arXiv:2206.12356},
year={2022}
}
@inproceedings{ramakrishnan2021hm3d,
title={Habitat-Matterport 3D Dataset ({HM}3D): 1000 Large-scale 3D Environments for Embodied {AI}},
author={Santhosh Kumar Ramakrishnan and Aaron Gokaslan and Erik Wijmans and Oleksandr Maksymets and Alexander Clegg and John M Turner and Eric Undersander and Wojciech Galuba and Andrew Westbury and Angel X Chang and Manolis Savva and Yili Zhao and Dhruv Batra},
booktitle={Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)},
year={2021},
url={https://openreview.net/forum?id=-v4OuqNs5P}
}
@article{collins2021abo,
title={ABO: Dataset and Benchmarks for Real-World 3D Object Understanding},
author={Collins, Jasmine and Goel, Shubham and Luthra, Achleshwar and
Xu, Leon and Deng, Kenan and Zhang, Xi and Yago Vicente, Tomas F and
Arora, Himanshu and Dideriksen, Thomas and Guillaumin, Matthieu and
Malik, Jitendra},
journal={arXiv preprint arXiv:2110.06199},
year={2021}
}