KPConv is a novel point convolution method that operates directly on point clouds without intermediate representations. It uses kernel points in Euclidean space to define convolution weights, allowing flexibility and adaptability to local geometry. Unlike fixed grid convolutions, KPConv can use any number of kernel points, enabling it to handle varying densities efficiently. The method also supports deformable convolutions, where kernel points are learned to adapt to local structures. KPConv is efficient and robust, outperforming state-of-the-art methods on various datasets for classification and segmentation tasks. It is also effective in handling complex and simpler tasks, with deformable KPConv excelling in challenging scenarios. The method uses a regular subsampling strategy and a regularization loss to ensure kernel points fit the point cloud geometry. Experiments show that KPConv achieves high performance on 3D shape classification, segmentation, and scene segmentation tasks. It outperforms other point convolution methods, especially in large and diverse datasets. The method is flexible and can be applied to various applications beyond classification and segmentation. KPConv is implemented with a modular architecture, including rigid and deformable versions, and is supported by extensive ablation studies and visualizations. The method is open-sourced for further research and development in point cloud convolutional architectures.KPConv is a novel point convolution method that operates directly on point clouds without intermediate representations. It uses kernel points in Euclidean space to define convolution weights, allowing flexibility and adaptability to local geometry. Unlike fixed grid convolutions, KPConv can use any number of kernel points, enabling it to handle varying densities efficiently. The method also supports deformable convolutions, where kernel points are learned to adapt to local structures. KPConv is efficient and robust, outperforming state-of-the-art methods on various datasets for classification and segmentation tasks. It is also effective in handling complex and simpler tasks, with deformable KPConv excelling in challenging scenarios. The method uses a regular subsampling strategy and a regularization loss to ensure kernel points fit the point cloud geometry. Experiments show that KPConv achieves high performance on 3D shape classification, segmentation, and scene segmentation tasks. It outperforms other point convolution methods, especially in large and diverse datasets. The method is flexible and can be applied to various applications beyond classification and segmentation. KPConv is implemented with a modular architecture, including rigid and deformable versions, and is supported by extensive ablation studies and visualizations. The method is open-sourced for further research and development in point cloud convolutional architectures.