An Improved Deep Multiple-input and Single-output PointNet for 3D Model Retrieval

PointNet extracts global shape features from the unordered point sets directly, respecting the permutation invariance of the input points; however, it fails to capture the fine-grained local shape features. In this paper, we extend PointNet to a multi-input and single-output structure by additionally feeding the scale-invariant heat kernel signature into PointNet to capture the fine-grained local shape features. To diversify the training data, we resample the points of each model randomly and generate a set of sub-samples, based on which PointNet calculates their classification scores. Then we adopt a plurality voting strategy to fuse the sub-sample level feature vectors to a model level descriptor, according to their classification scores. The experimental results demonstrate our proposed method outperforms the state-of-the-art retrieval methods on two 3D model benchmarks.