3D-Selfcutmix: Self-Supervised Learning For 3D Point Cloud Analysis

This paper extends Hadamard-coded supervised discrete hashing on real domain (R-HCSDH) using real-valued kernel transformations (RKTs) to one on complex/quaternion domain (C-HCSDH/H-HCSDH) using complex/quaternion KT (CKT/HKT). Supervised discrete hashing has recently attracted for its efficiency in data retrieval. Efficient learning of a hashing function is at the core of SDH. HCSDH simplifies its learning process by introducing Hadamard codes and shows its efficiency. Although many studies on SDH, including HCSDH, focus on the hashing function learning, KT, which is an initial step of SDH to generate a feature vector and has received less attention, also affects the performance and this motivates us to address in this work. Since conventional KTs are real-valued functions that only consider the distance between the input data and each anchor chosen from a training data set, it can not distinguish two anchors being equidistant. To solve this problem, we introduce CKT/HKT to consider not only the distance but also the angle between the input data and each anchor. Moreover, under theCKT and theHKT, we verify that Hadamard codes are still optimal for the HCSDH model. Experimental results show C-HCSDH and H-HCSDH outperform R-HCSDH in (cross-modal) data retrieval.