Category-Specific Progressive Refinement for Multimodal 3D Object Detection
Residual errors in multimodal 3D object detection can arise at different stages and vary across object categories. We propose a category-specific progressive refinement framework for a CenterHead–LoGoHead detector using RGB, thermal, and LiDAR inputs. Pedestrian Proposal Rescue (PPR) supplements first-stage pedestrian candidates through constrained proposal merging. Hard-Mismatch Geometry Refinement (HMR) predicts seven-dimensional residual corrections for pedestrian boxes using supervision segmented by proposal IoU. Scale-Routed Pyramid RoI Refinement (SPR) enhances Cyclist and Vehicle region-of-interest (RoI) features by combining multiple aggregation resolutions within a fixed physical region, with scale weights conditioned on category, distance, and response occupancy. Experiments use the 5637-frame DSERT-RoLL validation split. The original progressive model achieves 69.90% mAP and 62.17% mAPH, exceeding the seed-666 Base detector by 6.25 and 4.03 percentage points, respectively. Across seeds 666, 123, and 2026, the IoU-extended progressive configuration achieves 69.23 ± 0.09% mAP and 61.76 ± 0.06% mAPH. Joint HMR and SPR optimization from the seed-2026 PPR checkpoint achieves 69.93% mAP and 62.35% mAPH. Standalone module ablations and analyses across weather and lighting conditions characterize the contributions and remaining weaknesses of the refinement pipeline. The results support category- and stage-specific refinement within an existing tri-modal detection architecture.
Authors
- Xiaowei Xie
- 杨宏志
- Yu Mei (ORCID: https://orcid.org/0009-0008-3126-2974)
- Shiwei Huang
- Guangyu Xu
- Yuanping Xia
Institutions
- East China University of Technology (CN)
Publication Details
- Journal
- Applied Sciences
- Published
- 2026-11-19
- DOI
- https://doi.org/10.3390/app162211467
- Primary Topic
- Advanced Neural Network Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00