SV-RCNN: Robust Structural-Encoded Voxel Features for 3D Object Detection on Sparse Point Clouds

3D object detection in LiDAR point clouds faces significant challenges in sparse point clouds and unfamiliar environments. Existing methods often rely on simple multi-layer perceptrons or normalization techniques, limiting robustness in real-world settings, particularly for small or irregular objects. This work presents <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Structural Voxel RCNN (SV-RCNN)</b>, a generalizable 3D object detector that encodes geometric structure within each voxel through a structural convolutional operation. SV-RCNN introduces a learnable receptive field that adapts to object scale and sparsity, enabling reliable detection under varying conditions. Experimental results across multiple datasets demonstrate that SV-RCNN improves cross-domain performance and maintains strong accuracy without requiring additional adaptation steps. These results highlight the importance of structural encoding in enhancing generalization for real-world applications.