A Hierarchical Attention Fused Descriptor for 3D Point Matching

Motivated by recent successes on learning 3D feature representations, we present a Siamese network to generate representative 3D descriptors for 3D point matching in point cloud registration. Our system, dubbed HAF-Net, consists of feature extraction module, hierarchical feature reweighting and recalibration module (HRR), as well as feature aggregation and compression module. The HRR module is proposed to adaptively integrate multi-level features through learning, acting as a hierarchical attention fusion mechanism. The learnable feature pooling technique VLAD is extended into our aggregation module, which is further utilized to extract principal components of features and compress them into a low dimensional feature vector. To train our model, we amass a large dataset for 3D point matching. The dataset is composed of matched and unmatched point block pairs, which are automatically searched from existing reconstruction datasets with known poses. The experiments demonstrate that the proposed HAF-Net not only outperforms other state-of-the-art approaches in 3D feature representation but also has a good generalization ability in various tasks and datasets.