Thomas Hugues ; Tsai Yao-Hung Hubert ; Barfoot Timothy D. ; Zhang Jian

Abstract

In the field of deep point cloud understanding, KPConv is a uniquearchitecture that uses kernel points to locate convolutional weights in space,instead of relying on Multi-Layer Perceptron (MLP) encodings. While itinitially achieved success, it has since been surpassed by recent MLP networksthat employ updated designs and training strategies. Building upon the kernelpoint principle, we present two novel designs: KPConvD (depthwise KPConv), alighter design that enables the use of deeper architectures, and KPConvX, aninnovative design that scales the depthwise convolutional weights of KPConvDwith kernel attention values. Using KPConvX with a modern architecture andtraining strategy, we are able to outperform current state-of-the-artapproaches on the ScanObjectNN, Scannetv2, and S3DIS datasets. We validate ourdesign choices through ablation studies and release our code and models.