Composite convolution: A flexible operator for deep learning on 3D point clouds

Alberto Floris, Luca Frittoli, Diego Carrera, Giacomo Boracchi

Abstract: Highlights•We introduce the composite layer, a flexible and general alternative to convolutional operators for 3D point clouds.•Our composite layer extracts and compresses the spatial information from the 3D coordinates of points before combining it with their feature vectors.•Our composite layers have the same learning capabilities as existing point-convolutional layers while being more flexible in terms of design and number of parameters.•Neural networks based on our composite layers achieve similar performance as deeper, residual architectures based on existing point-convolutional layers in classification, segmentation, and anomaly detection.

External IDs:dblp:journals/pr/FlorisFCB24