Skip to content

Latest commit

 

History

History
90 lines (61 loc) · 3.66 KB

README.md

File metadata and controls

90 lines (61 loc) · 3.66 KB
Raw

SL-Cutscenes

sl-cutscenes is an easy-to-use extension framework for stillleben that generates realistic and visually diverse indoor scenes with physically interacting objects. With the help of stillleben, these scenes come in hi-res RGBD (stereo) frame sequences with dense semantic annotations (object classes/poses, instances, camera information, ...). This way, creating visually diverse video datasets for Computer Vision and Robotics becomes a piece of cake!

Examples

A collapsing stack of objects

image1

Household items falling onto a pool table

image2

Items about to be swept away by a robotic arm

image3

Fruits falling into a big bowl

image4

Bowling ball crashing into wooden blocks

image5

Balls falling into a laundry basket

image6

Installation

stillleben needs a custom installation due to special package requirements, so sl-cutscenes needs the following prerequisites:

  • python>3.6
  • conda
  • bash

For installation, executing the following steps:

  1. Create a new conda environment with python>3.6 and switch to the new environment.
  2. Install stillleben as shown here.
  3. Clone this repo to wherever you want and cd into it.
  4. Download the external asset data from here and unpack it into sl_cutscenes/assets/external_data.
  5. Install the remaining dependencies and the package: pip install ..

Usage

sl-cutscenes provides access to a variety of so-called "scenarios". These scenarios are object setups that lead to different physical interactions between them, sl-cutscenes comes with a wide variety of configuration options for the scenarios.

Generating scenes is done by running main.py, e.g.:

  • python main.py --scenario all --cameras 2 simulates each scenario once and renders the annotated video from 2 camera perspectives.
  • python main.py --scenario bowl --frames 90 --assemble-rgb simulates the bowl scenario once and until 90 frames have been produced, and additionally creates a video file from the rendered frames.
  • python main.py --scenario throw --iterations 3 --coplanar-stereo --sim-steps-per-frame 10 simulates the throw scenario three times with half the number of steps per frame (resulting in doubled fps) and captures it with a coplanar stereo camera.
  • python main.py -h Shows you the detailed argparse description of the different configuration options that can be controlled with optional arguments.

The generated data will be available in a time-stamped subfolder of the out directory of the repository.

Acknowledgements

  • The folder containing the object and texture data (downloadable from here) also contains an ACKNOWLEDGEMENT file for all acknowledgements regarding the used assets.
  • We'd like to thank Max Schwarz for insights and supportive development on stillleben to make this framework happen.

Citing

Please consider citing if you find our repository helpful (see "Cite this repository" in the repository's about section on github.)

License

This project is subject to an MIT License, see the LICENSE file.