A Novel Encoding Method for Pillar Networks: Dynamic Region and Height Fusion Approach

Methods based on the pillar framework for 3D object detection are well-known for their high computational efficiency. However, these approaches frequently encounter limitations, including insufficient connections between pillars, reduced responsiveness to variations in height, and restricted receptive fields, which collectively hinder their effectiveness, especially in detecting small-scale objects. To overcome these issues, we introduce a novel feature fusion strategy based on regional aggregation to improve pillar-based models. This method combines features from pillars located within specified regions while incorporating information about height variations, thereby enhancing the network's capability to perceive vertical spatial characteristics. The refined pillar features are subsequently processed through the feature extraction module and detection head, resulting in better object detection outcomes. Extensive evaluations using various benchmark datasets reveal that our approach significantly boosts the accuracy of detecting smaller objects while retaining its computational efficiency.