This repo provides some useful tools for using the HouseCat6D dataset [CVPR 2024 Highlight].
conda create -n housecat python=3.8
conda activate housecat
pip install -r requirement.txt
Copy the folder visualization
to the downloaded HouseCat6D folder.
One can visualize both rendered mask and 3D bounding box on entire frame given scene name and dataset directory.
Make sure to put obj_models_small_size_final
folder next to scene folders.
cd path/to/visualization
python vis_obj.py (path/to/dataset) (scene_name)
For example if the script is located into dataset folder and want to visualize scene01, simply run
python vis_obj.py ./ scene01
We support two ways of visualization. 3D will give an instant pyrender visualization, and 2D will save the image with the grasps rendered on the image plane. You can choose whether you want to downsample the visualized grasps.
cd path/to/visualization
python vis_grasp.py --split train --scene 1 --dimentional 3D --ds
We provide the reimplementation of VI-Net to show how to use HouseCat6D.
conda create -n vi-net python=3.9
conda activate vi-net
We tried with PyTorch 1.9 and CUDA 11.1.
cd VI-Net/lib/pointnet2/
pip install .
cd ../sphericalmap_utils/
pip install .
pip install gorilla-core==0.2.6.0
pip install opencv-python
pip install gpustat==1.0.0
pip install --upgrade protobuf
pip install scipy
Note that the latest gorilla-core
would fail. Modify path/to/HouseCat6D
in the VI-Net/config/housecat.yaml
.
Training with RGB-D
cd VI-Net
python train_housecat.py --gpus 0 --dataset housecat --mode r --config config/housecat.yaml
Training with RGB+P-D
python train_pol.py --gpus 0 --dataset housecat --mode r --config config/housecat.yaml
Training with RGB-D
cd VI-Net
python train_housecat.py --gpus 0 --dataset housecat --mode ts --config config/housecat.yaml
Training with RGB+P-D
cd VI-Net
python train_pol.py --gpus 0 --dataset housecat --mode ts --config config/housecat.yaml
HouseCat6D is released under CC BY 4.0.