Abstract
LiDAR-based 3D object detection is a cornerstone of autonomous driving, yet standard voxel-based methods often suffer from significant spatial information loss during down-sampling. This paper introduces Re-VoxelDet, a high-performance detector that addresses these limitations through a redesigned architecture. Re-VoxelDet incorporates a Multi-view Voxel Backbone (MVBackbone) for efficient feature extraction, a Hierarchical Voxel-guided Auxiliary Neck (HVANeck) to preserve fine-grained details, and a Rotation-based Group Head (RGHead) for specialized object localization. Experimental results on the Argoverse2 and nuScenes datasets demonstrate that Re-VoxelDet achieves state-of-the-art accuracy while running up to 35% faster than existing high-performance models.
Introduction
In autonomous driving, LiDAR sensors generate sparse point clouds that represent the 3D environment. To process this data, modern detectors typically discretize point clouds into voxels or pillars. While computationally efficient, the progressive down-sampling required by deep neural networks inevitably destroys fine-grained spatial information. This loss is particularly detrimental for detecting small objects, such as pedestrians, or accurately estimating the orientation of elongated objects like buses. Existing "neck" and "head" architectures often fail to recover this lost detail without incurring prohibitive computational costs. Re-VoxelDet is proposed to bridge this gap, offering a balanced solution for high-accuracy, real-time 3D detection.
Methodology
Re-VoxelDet revamps the standard 3D detection pipeline with three primary components designed to maximize information retention and inference efficiency.
Multi-view Voxel Backbone (MVBackbone)
The MVBackbone utilizes Sparse Reparameterized Feature (SRF) blocks. During the training phase, the network employs three parallel branches to capture features from multiple perspectives. At inference time, these branches are mathematically collapsed into a single, streamlined branch through structural reparameterization. This approach provides significant performance gains without increasing the computational overhead during deployment.
Hierarchical Voxel-guided Auxiliary Neck (HVANeck)
To combat the "wash-out" effect of deep layers on small objects, the HVANeck integrates features from earlier, higher-resolution stages of the backbone. It employs a novel Depth-wise Sparse Max-Pooling (SPMaxPool) mechanism to fuse high-resolution spatial details with high-level semantic information, ensuring that the geometric structure of distant or small objects is preserved.
Rotation-based Group Head (RGHead)
The RGHead categorizes objects into two distinct groups based on their geometric properties. One group focuses on "vehicle-like" objects with high aspect ratios, utilizing a specialized rotation-aware loss to improve heading estimation. The second group handles more isotropic objects like pedestrians. This task-specific specialization enhances both localization precision and processing speed.
Experimental Results
The performance of Re-VoxelDet was evaluated on the Argoverse2, Waymo Open Dataset (WOD), and nuScenes benchmarks.
Quantitative Performance
On the Argoverse2 validation set, Re-VoxelDet achieved 38.2 mAP, a substantial improvement over established baselines like CenterPoint (29.5 mAP) and VoxelNeXt (30.5 mAP). Furthermore, on the nuScenes dataset, the model demonstrated exceptional efficiency with an inference time of 81ms, making it approximately 35% faster than PillarNet (126ms).
Qualitative Analysis
Qualitative comparisons highlight the model's robustness in complex scenarios. As shown in Figure 1, Re-VoxelDet significantly reduces missed detections for small and distant objects, such as traffic cones and pedestrians, compared to standard voxel-based approaches.
Conclusion
Re-VoxelDet demonstrates that rethinking the neck and head architectures of 3D detectors can yield significant improvements in both accuracy and speed. By leveraging structural reparameterization and hierarchical feature fusion, the model effectively mitigates the spatial information loss inherent in voxel-based processing. Future work will explore the integration of multi-modal data, such as camera-LiDAR fusion, to further enhance perception in adverse environmental conditions.
Summary updated