Voxel-FPN: multi-scale voxel feature aggregation in 3D object detection from point clouds

Provided with extra depth information from 3D point cloud though, the difference of data modality between 3D point clouds and 2D RGB images makes it a big challenge in directly transplanting 2D detection techniques. Moreover, with the increase of dimensions and degrees-of-freedom, the objective of predicting exact position, size and orientation in 3D space requires highly-demanding efforts. In autonomous driving applications, RGB images and 3D point clouds could be simultaneously captured by camera and LIDAR sensors. Using either or both of two modalities, researchers explore effective and reliable solutions for 3D object detection tasks. In terms of representation learning, stateof-the-art work of 3D object detection could be divided into three kinds of methodology in whole: (a) fusion-based approaches, which synchronously fuse region features from RGB images and preprocessed 3D point clouds [7-9]; (b) 2D-detection-driven measures, to conduct subsequent object search in 3D subspace extended from 2D bounding boxes of detection results in RGB images [10]; (c) point-cloud-based methods, exploring the features and inner topology of points to detect 3D objects[11-19].