A Novel Adaptive Graph Transformer For Point Cloud Object Detection

With the advancement of autonomous driving technology, Lidar-driven 3D object detection has become a focal point of research. However, many existing point-driven approaches face challenges due to rigid local dependencies and entangled foreground-background points. The direct utilization of point clouds for 3D object detection is also complicated by the inherent sparsity and irregularities of Lidar data. To counter these challenges, we propose AGT-SSD a single-stage graph-based 3D detector to improve object performance especially for small objects. Which comprises two critical modules adaptive graph feature convolution module (AGF) and a point cloud transformer layer with self-attention mechanism (TSA). AGF module utilizes an adaptive graph kernel to process graph features, facilitating feature interactions within the feature domain to help detect small objects. Non-local transformer layer incorporates self-attention module, designed to capture long-range dependencies to encode global sematic and spatial feature to improve detection performance. This enhancement is attributed to the geometric structure refinement through graph convolution methods and the use of a self-attention mechanism to grasp long-range dependencies. Both modules employ non-local operations, adapt at discerning long-range correlations. Extensive experiments on the KITTI dataset demonstrated that our detector is not flexible and plug and play for other detectors but also highlight the superior accuracy of our method compared to other state-of-the-art detectors.