FG-Pillar: enhancing LiDAR-based small object detection via fine-grained pillar features

As a critical sensor for three-dimensional object detection, LiDAR enables the direct acquisition of the 3D coordinates of objects. However, the point clouds captured by LiDAR often exhibit low density, resulting in sparse information and a tendency to lose fine-grained details during feature extraction. To address these challenges, this study proposes a novel, to our knowledge, pillar-based network for 3D object detection. Specifically, a dual-stream adaptive feature enhancement module is introduced, which employs residual-gated feature refinement and a channel grouping strategy to achieve weighted feature fusion, thereby enhancing sensitivity to both fine-grained and global information. Additionally, a context-sensitive feature fusion module is designed, consisting of two complementary branches: the context aggregation branch, which aggregates global semantic information, and the semantic refinement branch, which strengthens local fine-grained interactions across channels. The collaborative operation of these two branches improves the sensitivity to sparse features and enhances multi-scale feature integration. On the KITTI dataset, the proposed network outperforms the baseline PillarNet, achieving improvements of 1.84% and 4.03% in <i>A</i><i>P</i>_3<i>D</i> for pedestrians and cyclists, respectively, as well as gains of 3.63% and 1.55% in <i>A</i><i>P</i>_{<i>B</i><i>E</i><i>V</i>}, and 1.55% and 2.51% in AOS. These experimental results demonstrate that the proposed network can effectively enhance detection accuracy and facilitate the broader application of LiDAR in small-object detection.