Learning general and distinctive 3D local deep descriptors for point cloud registration

An effective 3D descriptor should be invariant to different geometrictransformations, such as scale and rotation, robust to occlusions and clutter,and capable of generalising to different application domains. We present asimple yet effective method to learn general and distinctive 3D localdescriptors that can be used to register point clouds that are captured indifferent domains. Point cloud patches are extracted, canonicalised withrespect to their local reference frame, and encoded into scale androtation-invariant compact descriptors by a deep neural network that isinvariant to permutations of the input points. This design is what enables ourdescriptors to generalise across domains. We evaluate and compare ourdescriptors with alternative handcrafted and deep learning-based descriptors onseveral indoor and outdoor datasets that are reconstructed by using both RGBDsensors and laser scanners. Our descriptors outperform most recent descriptorsby a large margin in terms of generalisation, and also become the state of theart in benchmarks where training and testing are performed in the same domain.