Furuya and Ohbuchi: Deep Aggregation of Local 3d Features

Aggregation of local features is a well-studied approach for image as well as 3D model retrieval (3DMR). A carefully designed local 3D geometric feature is able to describe detailed local geometry of 3D model, often with invariance to geometric transformations that include 3D rotation of local 3D regions. For efficient 3DMR, these local features are aggregated into a feature per 3D model. A recent alternative, end-toend 3D Deep Convolutional Neural Network (3D-DCNN) [7][33], has achieved accuracy superior to the abovementioned aggregation-of-local-features approach. However, current 3D-DCNN based methods have weaknesses; they lack invariance against 3D rotation, and they often miss detailed geometrical features due to their quantization of shapes into coarse voxels in applying 3D-DCNN. In this paper, we propose a novel deep neural network for 3DMR called Deep Local feature Aggregation Network (DLAN) that combines extraction of rotation-invariant 3D local features and their aggregation in a single deep architecture. The DLAN describes local 3D regions of a 3D model by using a set of 3D geometric features invariant to local rotation. The DLAN then aggregates the set of features into a (global) rotation-invariant and compact feature per 3D model. Experimental evaluation shows that the DLAN outperforms the existing deep learning-based 3DMR algorithms.