Groupfree++: An Efficient 3D Object Detection Framework with Hybrid Sampling and Sparse Cross-Attentionx

Despite the inherent advantages of Transformer architectures in processing unordered point sets, their application in 3D object detection for autonomous driving remains limited due to computational complexity and information loss during point cloud downsampling. We present GroupFree++, a point-based 3D object detector that addresses these challenges through two key innovations: a hybrid sampling strategy that preserves crucial foreground points and a sparse cross-attention mechanism that significantly reduces computational overhead. Experiments on the KITTI benchmark demonstrate that our model achieves promising performance with an 8.84 % improvement in mean AP across categories while maintaining real-time inference speed at 17.56 FPS, making it practical for autonomous driving applications.