VoxT-GNN: A 3D object detection approach from point cloud based on voxel-level transformer and graph neural network

Abstract

Recently, a variety of LiDAR-based methods for the 3D detection of single-class objects, large objects, or in straightforward scenes have exhibited competitive performance. However, their detection performance in complex scenarios with multi - sized and multi - class objects is limited. We observe that the core problem leading to this phenomenon is the insufficient feature learning of small objects in point clouds, making it difficult to obtain more discriminative features. To address this challenge, we propose a 3D object detection framework based on point clouds that takes into account the detection of small objects, termed VoxT-GNN. The framework comprises two core components: a Voxel-Level Transformer (VoxelFormer) for local feature learning and a Graph Neural Network Feed-Forward Network (GnnFFN) for global feature learning. By embedding GnnFFN as an intermediate layer between the encoder and decoder of VoxelFormer, we achieve flexible scaling of the global receptive field while maximally preserving the original point cloud structure. This design enables effective adaptation to objects of varying sizes and categories, providing a viable solution for detection applications across diverse scenarios. Extensive experiments on KITTI and Waymo Open Dataset (WOD) demonstrate the strong competitiveness of our method, particularly showing significant improvements in small object detection. Notably, our approach achieves the second-highest mAP of 65.44% across three categories (car, pedestrian, and cyclist) on KITTI benchmark. The source code is available at https://github.com/yujianxinnian/VoxT-GNN.