RePack: Dense Object Packing Using Deep CNN with Reinforcement Learning

Abstract

We propose a new object packing approach, RePack, to arrange a series of identical image objects to a rectangular canvas densely by a deep CNN with reinforcement learning. In our approach, adding a new object to an image pack of existing objects is modeled as classification of possible pack configurations by a CNN. To iteratively reinforce the CNN, pack trees are built to identify object overlaps and to find denser pack configurations for reinforcement training. Such a reinforcement learning process for enhancing a CNN for dense object packing is rarely seen in previous literature. Preliminary experimental results show that the reinforced deep CNN can generate dense object packs in a sequential manner for circular, triangular and quadrilateral objects.