Local grid rendering networks for 3D object detection in point clouds

Recognizing local geometric patterns in point clouds is crucial for accurate 3D object detection. However, existing point-based methods often fail to capture sufficiently detailed local geometry due to their limited representation capability. Conversely, CNNs can effectively model local geometry but become computationally intensive when processing densely voxelized grids derived from point clouds. To address this challenge, we propose the Local Grid Rendering Network (LGR-Net), an efficient backbone architecture designed specifically to exploit CNNs’ representational strength while maintaining computational efficiency by leveraging the inherent sparsity of point cloud data. LGR-Net hierarchically samples key local regions, rendering them into sparse, low-resolution grids via a Local Grid Rendering operation. These grids are then processed with a lightweight CNN to effectively capture detailed geometric information at minimal computational cost. To further align our method with the sparse geometric nature of real-world point cloud objects, we introduce a novel density-aware copy-paste data augmentation strategy, specifically designed to enhance the learning of detailed geometric features by generating realistic, sparsely populated training samples. Extensive experiments demonstrate that our approach outperforms strong baselines in both indoor and outdoor 3D object detection scenarios, with minimal additional computational overhead.