EPDet: Enhancing point clouds features with effective representation for 3D object detection

Le résumé fourni par la source

Effective 3D object detection relies on strong feature representation. Both global features and representative characteristics of the objects are vital for detection. However, traditional convolution’s perception range limits large receptive fields, and current object feature representation has room for enhancement. Addressing these concerns, we introduce a 3D outdoor object detector to enhance the point clouds feature, referred to as EPDet. Specifically, for global features, we propose a BEV-Offset Transformer in the BEV (Bird’s Eye View) domain. This adaptable module enhances semantic connections among objects, suiting various 3D detection methods. In addition, to refine point cloud features, we employ Focal Conv as our 3D ( 3-dimensional) backbone, exploring multi-modal fusion effects. In the 2D (2-dimensional) backbone, our Pyramid-like Conv captures detailed contextual features. EPDet performs well in the dense object scene owing to scene-wide global features captured by BEV Offset Transformer. In the multi-class tests, EPDet excels in detecting smaller objects due to more refined features represented by Focal Conv and Pyramid-like Conv. In experiments, as a plug-and-play module, we validate the BEV Offset Transformer’s effectiveness across single-stage (SECOND), two-stage (Voxel-RCNN), and multi-stage (CasA) algorithms. Robustness is tested on KITTI, NuScenes, and ONCE datasets. The proposed EPDet, in the KITTI subset, EPDet (CasA-based) achieves an impressive 85.56% accuracy in the car category. EPDet (Voxel-RCNN-based) surpassed baseline 1.65% mAP (mean Average Precision) (moderate subsets) in multi-class detection. The precision of EPDet is on par with the SotA (State of the Art) 3D outdoor object detectors based on point clouds.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.