A Hierarchical Segmentation Approach with Convolution-Recursive Deep Learning for 3D Multi-Object Recognition under Partial Occlusion Conditions

Depth data based object recognition has recently emerged as a challenging research topic. In this work, we develop a novel approach to perform detection and recognition of occluded 3D objects. We propose a hierarchical segmentation algorithm in order to obtain the homogeneous sub-regions contained in each depth frame which in turn facilitates the recognition under severe occlusion conditions. Our proposal consists of three steps: the first step is to build a tree structure contains all key sub-surfaces visible in the depth frame with their intra-hierarchical relations. Thereafter, we draw a classification prediction for all nodes based on a combination of convolution and recursive neural networks. Finally, we employ the hierarchy scheme to refine the classification results. Our proposal obtained competitive results and proved to be invariant to objects scale, rotation, and viewpoint variations.