In the field of deep point cloud understanding, KPConv is a unique architecture that uses kernel points to localize convolutional weights in space, instead of relying on multi-layer perceptron (MLP) encodings. While initially achieving success, it has since been surpassed by recent MLP networks employing updated training designs and strategies. Based on the kernel point principle, we present two novel designs: KPConvD (KPConv in depth), a lighter design that allows the use of deeper architectures, and KPConvX, a novel design that scales the depth-wise convolutional weights of KPConvD with core attention values. By using KPConvX with a modern architecture and training strategy, we can outperform current state-of-the-art approaches on the ScanObjectNN, Scannetv2, and S3DIS datasets. We validated our design choices through ablation studies and published our code and models.
