DSA3D: Dynamic Salience-Aware Feature Learning for 3D Object Detection

Accurate 3D object detection utilizing sparse LiDAR point clouds stands as a critical bottleneck in autonomous perception, largely stemming from the data's intrinsic irregularity and varying spatial density. Existing voxel-based paradigms often suffer from feature quantization loss, where heuristic pooling operations discard critical micro-geometric details, while point-based methods incur prohibitive computational costs. Toward this end, we propose DSA3D, an integrated framework that synergizes hierarchical feature abstraction with dynamic computation.Specifically, we introduce a Hierarchical Voxel Feature Encoder incorporating the Spatial-Voxel Attention (SVA) mechanism. By leveraging a dual-stage residual design, SVA explicitly models intra-voxel geometric distributions and inter-channel semantic dependencies, effectively mitigating quantization artifacts. Furthermore, we design a Spatially-Adaptive 3D Backbone equipped with Dynamic Sparse Convolution (DSC). This operator functions as a learnable importance evaluator, adaptively allocating computational budget to semantically salient regions while pruning redundant background noise. Finally, we present the Spatial Context Aggregation Network (SCA-Net) to facilitate wide-area reasoning in the BEV domain. Thorough evaluations on the KITTI benchmark reveal that DSA3D attains highly competitive performance, yielding 82.53% AP for the Car category.