CHOP is the core vision algorithm in PaCMan EU project,
which focuses on building robots that will load a dishwasher.
In order for the robots to interact with our complex world, we need successful computer vision systems. Recent advances in image categorization and detection challenges suggest that deep, hierarchical learning methods can capture highly implicit visual patterns. CHOP is a learning algorithm that builds part-based object models across multiple scales. It is built upon the principles of LHOP by Leonardis et al., as well as many other hierarchical compositional methods. At the core of our algorithm is a graph mining method, SUBDUE, which implements a Minimum Description Length based part selection criterion.
Paper
We have presented an early version of CHOP at ECCV 2014, in Zurich, Switzerland in September 2014. You can download the paper using this link, while the supplementary meterial can be found here.
Code
A MATLAB implementation of CHOP has been released in September, 2014. It builds upon our ECCV paper and incorporates additional features, such as grouping compositions based on their geometric/appearance similarities.