EP-Net: Improving Point Cloud Learning Efficiency Through Feature Decoupling

Point cloud analysis has been significantly influenced by the PointNet family, which has achieved remarkable accuracy across various benchmarks. However, enhancing the efficiency of these models without compromising performance remains a challenging task. In this article, we investigate the efficiency of the state-of-the-art (SOTA) point-based method PointNeXt without degrading its performance and propose a more efficient and powerful network named efficient PointNet (EP-Net). In particular, we propose a decoupled feature aggregation (DFA) module that decouples the learning of geometric and semantic features to capture more discriminative information, rather than the vanilla set abstraction (SA) module aggregating them for coarse joint processing. Besides, we develop a more effective sampling strategy for EP-Net, compressing processing complexity from $\mathcal {O}(n \log (n))$ to linear. It selects points nearest to the center of each nonempty grid cell as the new sampling points. Impressively, we achieve a dual increase in both speed and accuracy, attributed to the more refined feature aggregation and more efficient sampling. Furthermore, we demonstrate the extensibility of our proposal toward other point cloud networks, from the perspective either of boost efficiency or performance prompt. Conditioned on extensive experiments on point cloud classification and segmentation tasks, EP-Net reports an SOTA performance with 73.3% mean IoU on S3DIS Area5, boosting their baseline PointNeXt by 2.5%, with only 10.6% FLOPs and 37% parameters. On extensibility experiments, our strategy boosts PointMLP with an improvement of 0.9% mAcc on the ScanObjectNN dataset.