Decoupled Local Aggregation for Point Cloud Learning

The unstructured nature of point clouds demands that local aggregation beadaptive to different local structures. Previous methods meet this byexplicitly embedding spatial relations into each aggregation process. Althoughthis coupled approach has been shown effective in generating clear semantics,aggregation can be greatly slowed down due to repeated relation learning andredundant computation to mix directional and point features. In this work, wepropose to decouple the explicit modelling of spatial relations from localaggregation. We theoretically prove that basic neighbor pooling operations cantoo function without loss of clarity in feature fusion, so long as essentialspatial information has been encoded in point features. As an instantiation ofdecoupled local aggregation, we present DeLA, a lightweight point network,where in each learning stage relative spatial encodings are first formed, andonly pointwise convolutions plus edge max-pooling are used for localaggregation then. Further, a regularization term is employed to reducepotential ambiguity through the prediction of relative coordinates.Conceptually simple though, experimental results on five classic benchmarksdemonstrate that DeLA achieves state-of-the-art performance with reduced orcomparable latency. Specifically, DeLA achieves over 90\% overall accuracy onScanObjectNN and 74\% mIoU on S3DIS Area 5. Our code is available athttps://github.com/Matrix-ASC/DeLA .