Robotic grasp dataset for multi-object multi-grasp evaluation with RGB-D data.
This dataset is annotated using the same protocal as Cornell Dataset, and can be used as multi-object extension of Cornell Dataset.
Download RGB-D data and put into grasp_multiObject/rgbd/
Each testing data has one RGB image (rgb_xxxx) and one depth image (depth_xxxx).
The corresponding grasp annotation (rgb_xxxx_annotations) can be found in grasp_multiObject/annotation/
mkdir rgd
run rgbd2rgd
you will have RG-D data in grasp_multiObject/rgd/
mkdir rgd_cropped320
mkdir rgb_cropped320
run image2txt
you will have cropped RGB and RGD images in grasp_multiObject/rgd_cropped320/
and grasp_multiObject/rgb_cropped320/
, respectively.
also, you will have corresponding annotation files, as well as a full list of image path.
run visualizationGripper
this file shows a simple example to visualize ground truth grasps
git clone https://github.com/ivalab/grasp_annotation_tool
you can annotate grasps on your own data with this simple tool!
Both dataset and annotation tool can also be found here
If you find it helpful for your research, please consider citing:
@inproceedings{chu2018deep,
title = {Real-World Multiobject, Multigrasp Detection},
author = {F. Chu and R. Xu and P. A. Vela},
journal = {IEEE Robotics and Automation Letters},
year = {2018},
volume = {3},
number = {4},
pages = {3355-3362},
DOI = {10.1109/LRA.2018.2852777},
ISSN = {2377-3766},
month = {Oct}
}
If you encounter any questions, please contact me at fujenchu[at]gatech[dot]edu