The HANDAL dataset is category-level dataset for object pose estimation and affordance prediction, with a focus on hardware and kitchen tool objects that are of the proper size and shape for functional robotic grasping (e.g. pliers, utensils, and screwdrivers). The dataset consists of 308k annotated image frames from 2.2k videos of 212 real-world objects in 17 categories.
Dataset is available for download from Google Drive. Dataset is released under CC-BY-NC-SA-4.0 license.
Dataset follows the BOP format, with some minor modifications:
- Image size is provided in
scene_camera.jsonfor each scene, since multiple camera sensors were used to capture data - Depth is rendered from NeRF reconstructions of the scene (Instant NGP) rather than from a depth sensor
- RGB images are compressed using JPEG rather than PNG
For more details, see the project page and our 2023 IROS paper.
Our annotation process is streamlined, requiring only a single off-the-shelf camera and semi-automated processing, allowing us to produce high-quality 3D annotations without crowd-sourcing. This section covers how to install the HANDAL toolkit and use it to annotate static scenes.
Our toolkit requires an NVIDIA GPU capable of running Instant-NGP. We have only tested the toolkit on Ubuntu.
We recommend installing the HANDAL pipeline in a separate Python environment, using Python 3.10 or greater.
conda create -n handal-toolkit python=3.10
git clone https://https://github.com/NVlabs/HANDAL.git
cd HANDAL
pip install -r requirements.txt
Our pipeline leverages the Instant-NGP Python bindings for scene reconstruction, depth estimation, and mesh construction.
Follow the instructions to build Instant-NGP and its Python bindings.
Some potential headaches may be avoided by building the Python bindings in the same Python environment as the one used to install the HANDAL toolkit.
Finally, set the CONFIG_NGP_PATH variable in configs/default_config.sh to the root directory of the Instant-NGP repository.
The following commands worked for our local build of Instant-NGP. We share them in case they are helpful, but we recommend following the official instructions if any problems are encountered:
conda activate handal-toolkit
cd HANDAL
git clone --recursive https://github.com/nvlabs/instant-ngp submodules/instant-ngp
cd submodules/instant-ngp
conda install -c conda-forge libstdcxx-ng=12 --yes # fix the missing GLIBCXX 3.4.30 error
pip install -r requirements.txt
cmake . -B build -DCMAKE_BUILD_TYPE=RelWithDebInfo
cmake --build build --config RelWithDebInfo -j
python ./scripts/run.py --help
To initialize the pipeline, we require rough segmentation masks of the object to be annotated. We recommend considering the video segmentation tool Cutie, a follow-up work to XMem. To avoid conflicting with the other requirements, we recommend installing it in a separate Python environment.
To test the installation, download the example scenes from Google Drive and extract them to example/. Object segmentation masks are already provided for these examples in the input_masks/ directory.
To generate the initial reconstruction and annotation of the reference scan, run the following command:
bash run.sh example/drill_reference
If the command runs successfully, you will find initial annotations and results in example/drill_reference/, including raw BOP annotations bop_raw/scene_*_initial.json, a mesh with vertex colors meshes/colored.ply, and rendered depth maps depth_scene/. The initial annotations lack ground truth scale because we use RGB-only inputs. The next steps will interactively place the reference object in a canonical pose and estimate the correct scale.
First, ensure that example/drill_reference/config.sh contains the following line to indicate that this is a reference scene:
CONFIG_IS_REFERENCE=1
Then, run the following command:
bash run.sh example/drill_reference --interactive
First, an interactive tool will place the object in a canonical pose. The canonical pose is automatically initialized using an oriented bounding cuboid fit to the object mesh.
Use the keys r, g, and b to rotate the object in 1 degree increments around the x, y, and z axes, respectively.
(The shift key can be used to rotate in the opposite direction, the ctrl key can be used to make 90 degree rotations, and the alt key can be used to make 0.1 degree rotations.)
Use the v key to cycle through canonical viewing directions (corresponding to the BOP documentation): front view (red x-axis pointing out of screen), side view (green y-axis pointing out of screen), and top view (blue z-axis pointing out of screen).
For this example, in the front view, we recommend rotating such that the top of the drill is aligned with the positive z axis (blue arrow) and the front of the drill is pointing in the negative y direction (green arrow).
Once the object is in the desired pose, press q to quit the interactive tool and write the pose to example/drill_reference/canonical_pose.json.
Next, you will be prompted to enter the measured dimensions of the object in millimeters. For the drill example, the dimensions are 230mm x 130mm x 75mm corresponding to the height, width, and depth of the object in BOP canonical pose. (Or, equivalently, the size of bounding cuboid along the z-, y-, and x-axes, respectively.)
The estimated scale will be written to example/drill_reference/canonical_scale.json.
Finally, the pipeline will apply the canonical pose and scale and finalize the BOP annotations in example/drill_reference/bop/.
To process an additional scene (e.g. example/drill_scene1) and align it to the reference scene, first ensure that example/drill_scene1/config.sh contains the following line to indicate the reference scene (relative to the scene's directory):
CONFIG_REFERENCE_SCENE=../drill_reference
Then run the following command:
bash run.sh example/drill_scene1 --interactive
After the initial reconstruction, the interactive GUI will first prompt for the canonical pose, followed by a prompt for scale. (Be careful when scaling as some dimensions may not be accurate if the object was occluded.) Then the interative GUI will guide the alignment of the object to the reference scan. In addition to the rotation keys used to set the canonical pose, the x, y, and z keys may be used to translate the object in 1mm increments along the x, y, and z axes. (As before, the shift, ctrl, and alt keys may be used to translate in the opposite direction, make 50mm adjustments, and make 0.2mm adjustments, respectively.)
Finally, the pipeline will use this transformation and the mesh copied from the reference scene to generate final annotation files in example/drill_scene1/bop/.
@InProceedings{handaliros23,
title={{HANDAL}: A Dataset of Real-World Manipulable Object Categories with Pose Annotations, Affordances, and Reconstructions},
author={Andrew Guo and Bowen Wen and Jianhe Yuan and Jonathan Tremblay and Stephen Tyree and Jeffrey Smith and Stan Birchfield},
booktitle={IROS},
year={2023}
}The dataset is released under CC-BY-NC-SA-4.0 license. The code is released under the NVIDIA Source Code License.
Copyright © 2023-2024, NVIDIA Corporation. All rights reserved.