simData_imgSaver

ros package for saving simulation images in Gazebo with Kinect

Get the package

cd ~/catkin_ws/src/
git clone https://github.com/ivalab/simData_imgSaver.git
cd ..
catkin_make

Clone the simData

cd ~/catkin_ws/src/
git clone https://github.com/ivalab/simData.git 

rename the /3Dwarehouse/tools to /warehouse/models

Preprocess data

in utils/move_model_files.m modify the dataRoot

dataRoot = '/home/YOUR/PATH/TO/catkin_ws/src/simData_v2/warehouse/models';

run move_model_files.m in MATLAB

Copy models

copy all models to /home/YOURNAME/.gazebo/models/

NOTE:
you can add export GAZEBO_RESOURCE_PATH=$GAZEBO_RESOURCE_PATH:/home/YOUR/PATH/TO/catkin_ws/src/simData_v2/warehouse/ in ~/.bashrc then you can might skip this step
but change line 34:

fprintf(fsdf, ['            <uri>file://models/' FolderName '/mesh/' filename '.dae</uri>\n']);

Open Gazebo

rosrun gazebo_ros gazebo

Run imgSaver with object

in visuomotor_grasp_3D_Box.py
modify MODEL_DIR to /home/YOUR/PATH/TO/catkin_ws/src/simData_v2/warehouse/models

python visuomotor_grasp_3D_Box.py

the files will be saved to /dataset

Run imgSaver with affordance

in visuomotor_grasp_3D_Box_affordance.py
modify MODEL_DIR to /home/YOUR/PATH/TO/catkin_ws/src/simData_v2/warehouse/models

python visuomotor_grasp_3D_Box_affordance.py

it loads corresponding affordance .dae files with dark table and highlight the affordance with specific color
the files will be saved to /dataset

Remove shadow effect

When saving affordance images, the light source causes shadow and affects the color. The light source is not loadable in code (at lease in older version of Gazebo). One quick and effective solution is to add multiple light sources in Gazebo with GUI.

Click on the parallel light source and select a location to place, and click on the icon again to turn it on. Also for tall object such as large mug, it is useful to also put several all-directional light source in front of the object.

Process real data

To generate bonding boxes for UMD dataset, put simData_imgSaver/utils/boundingBoxGenerator.m in YOUR/PATH/TO/UMD_affordance/part-affordance-dataset/ and run the script.

We segment the blue lazy Susan table and binarize the output to get the bounding boxes in HSV space. You can modify the value to adjust to your case.

hsv(:,:,3) > INTENSITY_T; % filter out black color  
hsv(:,:,2) > 0.33; % filter out white color 

Real data with bounding box

Sim data

Sim data with bounging box

Sim data with grasp affordance

Sim data with hit affordance

Citation

If you find it helpful for your research, please consider citing:

@inproceedings{chu2019learning,
  title = {Learning Affordance Segmentation for Real-world Robotic Manipulation via Synthetic Images},
  author = {F. Chu and R. Xu and P. A. Vela},
  journal = {IEEE Robotics and Automation Letters},
  year = {2019},
  volume = {4},
  number = {2},
  pages = {1140-1147},
  DOI = {10.1109/LRA.2019.2894439},
  ISSN = {2377-3766},
  month = {April}
}

Contact

If you encounter any questions, please contact me at fujenchu[at]gatech[dot]edu