cmitash / Physim Dataset Generator

PHYSIM-DATASET-GENERATOR

This repository implements a software tool for synthesizing images of physically realistic cluttered scenes using 3D CAD models as described in our paper:

A Self-supervised Learning System for Object Detection using Physics Simulation and Multi-view Pose Estimation (pdf)(website)

By Chaitanya Mitash, Kostas Bekris, Abdeslam Boularias (Rutgers University).

To appear at the IEEE International Conference on Intelligent Robots and Systems (IROS), Vancouver, Canada, 2017.

Citing

To cite the work:

@article{physim,
  title={A Self-supervised Learning System for Object Detection using Physics Simulation and Multi-view Pose Estimation},
  author={Mitash, Chaitanya and Bekris, Kostas and Boularias, Abdeslam},
  journal={arXiv:1703.03347},
  year={2017}
}

Setup

1. Download and extract Blender
2. In ~/.bashrc, add line export BLENDER_PATH=/path/to/blender/blender
3. Get the get-pip.py file from the pip documentation
4. Install the yaml package in the python packaged with blender using the commands below.

$ /path/to/blender/2.xx/python/bin/python3.5m /path/to/get-pip/get-pip.py
$ /path/to/blender/2.xx/python/bin/python3.5m /blender-version/2.xx/python/bin/pip install pyyaml

Demo

1. In ~/.bashrc, add line export PHYSIM_GENDATA=/path/to/repo.
2. Rename config.yml.shelf or config.yml.shelf to config.yml and modify simulation parameters if required.
3. Run python generate_pictures.py
4. The generated data can be found in the folder rendered_images. Available environments are table and shelf.

Output

1. Images of scenes.
2. Labeled bounding box files for each scene <label, tl_x, tl_y, br_x, br_y> or if the pixel label mode is selected, a pixel-wise labeled image is generated for each scene where the pixel value is the ground-truth class value.
3. Debug images indicating the bounding-boxes over the objects.
4. .blend files to debug the simulation parameters.

Parameters

the example cfg files contain the parameters of simulation.

camera:
  num_poses: <number of views to render from>
  camera_poses: [[pos_x, pos_y, pos_z, quat_w, quat_x, quat_y, quat_z], ...]
  camera_intrinsics: [[f_x, 0.0, c_x],[0.0, f_y, c_y],[0.0, 0.0, 1.0]]

rest_surface:
  type: shelf
  surface_pose: [pos_x, pos_y, pos_z, quat_w, quat_x, quat_y, quat_z]

Models: [model_1, model_2, ...]

params:
  num_images: <number of training images>
  label_type: box
  minimum_objects_in_scene: <minimum object per scene>
  maximum_objects_in_scene: <minimum object per scene>
  range_x: [<min_x>, <max_x>]
  range_y: [<min_y>, <max_y>]
  range_z: [<min_z>, <max_z>]
  num_simulation_steps: <number os simulation steps to run>
  light_position_range_x: [<min_x>, <max_x>]
  light_position_range_y: [<min_y>, <max_y>]
  light_position_range_z: [<min_z>, <max_z>]