LoGoNet: A Local-to-Global Fusion Network for 3D Object Detection
LoGoNet: A Local-to-Global Fusion Network for 3D Object Detection
LiDAR-camera fusion methods have demonstrated impressive performance in 3D object detection. However, current multi-modal methods primarily perform global fusion, lacking fine-grained region-level information and resulting in suboptimal fusion performance.
This paper presents LoGoNet, a novel Local-to-Global fusion network, for LiDAR-camera fusion in 3D object detection. LoGoNet performs fusion at both local and global levels to capture comprehensive information for accurate object detection.
Global Fusion (GoF)
LoGoNet's GoF builds upon previous research but utilizes point centroids to represent voxel features more precisely, enabling better cross-modal alignment between LiDAR and camera data.
Local Fusion (LoF)
For LoF, each proposal is divided into uniform grids, and grid centers are projected onto the image plane. Image features around these projected points are sampled and fused with position-decorated point cloud features. This approach maximizes the utilization of rich contextual information surrounding the proposals.
Feature Dynamic Aggregation (FDA)
The Feature Dynamic Aggregation (FDA) module facilitates information interaction between locally and globally fused features, producing more informative multi-modal features for improved detection accuracy.
Experimental Results
Extensive experiments on the Waymo Open Dataset (WOD) and KITTI datasets demonstrate that LoGoNet outperforms all state-of-the-art 3D detection methods. Notably, LoGoNet achieves the following:
- Ranks 1st on the Waymo 3D object detection leaderboard.* Achieves 81.02 mAPH (L2) detection performance.* Surpasses 80 APH (L2) on three classes simultaneously for the first time.
Conclusion
LoGoNet presents a novel approach to LiDAR-camera fusion for 3D object detection. Its local-to-global fusion strategy, coupled with point centroid representation and the FDA module, leads to significant performance improvements.
Code is available at https://github.com/sankin97/LoGoNet.
原文地址: https://www.cveoy.top/t/topic/Uxx 著作权归作者所有。请勿转载和采集!