This repo contains code to process and render images of the MIT Pushing dataset (http://web.mit.edu/mcube//push-dataset/).

Preprocessing

First, create a folder structure like [pushdata folder]/[surface name]/ by downloading the zip-files for each surface to [pushdata folder]/ and unzipping them there. This should result in a folder structure like this: [pushdata folder]/[surface name]/[object name]/[object_name]_h5.zip

Below we will use ~/pd as the pushdata root folder. For example, the data of rect1 on ABS surface will be at ~/pd/abs/rect1/rect1_h5.zip.

First preprocess the data using

scripts/preprocess.py --source-dir ~/pd

This will

• remove redundant data entries
• treat some (not all) of the jumps in object orientation
• transform the orientation to [-pi, pi]
• synchronize the data by resampling to a given frequency
• set the initial object position and orientation to zero
• add information about the push (angle, velocity...) to the h5 data files

This will save all the preprocessed h5 files for object [object_name] on surface [surface_name] to ~/pd/[surface name]/[object name]. If you want to write output to another directory than the source directory (~/pd), you can specify this with the --out-dir argument.

RGB-D rendering

The preprocessing has to be done before.

To render RGB-D pictures from the preprocessed images, refer to scripts/render_scene.py.

Assuming we are at the top folder of pdproc, a rendering demo can be run by

scripts/render_scene.py --source-folder ~/pd --out-dir .

This will generate a series of images in jpeg in the current folder.

Contact annotation

To annotate the preprocessed h5 files with contact points and surface normals, please refer to scripts/render_scene.py.