FCAF3D: Fully Convolutional Anchor-Free 3D Object Detection

Recently, promising applications in robotics and augmented reality have attracted considerable attention to 3D object detection from point clouds. In this paper, we present FCAF3D - a first-in-class fully convolutional anchor-free indoor method. It is simple yet effective method that uses voxel representation of cloud processes voxels with sparse convolutions. can handle large-scale scenes minimal runtime through single feed-forward pass. Existing methods make prior assumptions on the geometry objects, argue it limits their generalization ability. To get rid any assumptions, propose novel parametrization oriented bounding boxes allows obtaining better results purely data-driven way. The proposed achieves state-of-the-art terms [email protected] ScanNet V2 (+4.5), SUN RGB-D (+3.5), S3DIS (+20.5) datasets. code models are available at https://github.com/samsunglabs/fcaf3d.