DeLA: An extremely faster network with decoupled local aggregation for large scale point cloud learning

Abstract

With advances in data collection technology, the volume of recent remote sensing point cloud datasets has grown significantly, posing substantial challenges for point cloud deep learning, particularly in neighborhood aggregation operations. Unlike simple pooling, neighborhood aggregation incorporates spatial relationships between points into the feature aggregation process, requiring repeated relationship learning and resulting in substantial computational redundancy. The exponential increase in data volume exacerbates this issue. To address this, we theoretically demonstrate that if basic spatial information is encoded in point features, simple pooling operations can effectively aggregate features. This means the spatial relationships can be extracted and integrated with other features during aggregation. Based on this concept, we propose a lightweight point network called DeLA (Decoupled Local Aggregation). DeLA separates the traditional neighborhood aggregation process into distinct spatial encoding and local aggregation operations, reducing the computational complexity by a factor of K, where K is the number of neighbors in the K-Nearest Neighbor algorithm (K-NN). Experimental results on five classic benchmarks show that DeLA achieves state-of-the-art performance with reduced or equivalent latency. Specifically, DeLA exceeds 90% overall accuracy on ScanObjectNN and 74% mIoU on S3DIS Area 5. Additionally, DeLA achieves state-of-the-art results on ScanNetV2 with only 20% of the parameters of equivalent models. Our code is available at https://github.com/Matrix-ASC/DeLA.