Research map / Detection and segmentation / 3D detection and driving
Deformable Feature Aggregation for Dynamic Multi-modal 3D Object Detection
Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinhong Jiang, Feng Zhao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Cross-Domain DeformCAFA with image branch (ablation) | CenterPoint · nuScenes val 1/8 | mAP: 50.28 -> 56.96 | Improved |
| add image-level dropout training (ablation) | CenterPoint + DeformCAFA · nuScenes val 1/8 | mAP: 56.96 -> 57.03 | No clear effect |
| add depth-aware GT-AUG (ablation) | CenterPoint + DeformCAFA + dropout · nuScenes val 1/8 | mAP: 57.03 -> 58.45 | Improved |
| use multiply of point and image features for query token instead of concat (ablation) | Cross-Domain DeformCAFA · nuScenes val 1/8 | mAP: 58.01 -> 58.45 | Improved |
| use both point and image features for token instead of points only (ablation) | Cross-Domain DeformCAFA · nuScenes val 1/8 | mAP: 57.10 -> 58.45 | Improved |
| replace AutoAlign fusion with DeformCAFA (ablation) | CenterPoint · nuScenes val 1/8 | mAP: 56.69 -> 58.45 | Improved |
| use depth-aware GT-AUG instead of MoCa GT-AUG (ablation) | CenterPoint · nuScenes val 1/8 | mAP: 53.08 -> 53.48 | Improved |
| mix-up ratio alpha 1.0 instead of 0.5 to 0.8 (ablation) | depth-aware GT-AUG · nuScenes val 1/8 | mAP: drops about 0.7 mAP | Hurt |
| train with 3 images instead of 1 (ablation) | image-level dropout · nuScenes val 1/8 | mAP: 57.93 -> 58.45 | Improved |
| train with 6 images instead of 3 (ablation) | image-level dropout · nuScenes val 1/8 | mAP: 58.45 -> 58.51 | No clear effect |
| infer with 3 images instead of 6 | AutoAlignV2 · nuScenes val 1/8 | mAP: 58.45 -> 54.32 | Hurt |
| infer with 0 images (LiDAR only) | AutoAlignV2 · nuScenes val 1/8 | mAP: 58.45 -> 50.29 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- AFDetV2 — AFDetV2: Rethinking the Necessity of the Second Stage for Object Detection from Point Clouds (AAAI 2022) · better on 1 of 1 shared result
- 3D-CVF — 3D-CVF: Generating Joint Camera and LiDAR Features Using Cross-View Spatial Feature Fusion for 3D Object Detection (ECCV 2020) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Homogeneous Multi-modal Feature Fusion and Interaction for 3D Object Detection (ECCV 2022 · similarity 0.90)
- CCF: Complementary Collaborative Fusion for Domain Generalized Multi-Modal 3D Object Detection (CVPR 2026 · similarity 0.88)
- MV2DFusion: Leveraging Modality-Specific Object Semantics for Multi-Modal 3D Detection (TPAMI 2025 · similarity 0.88)
- STXD: Structural and Temporal Cross-Modal Distillation for Multi-View 3D Object Detection (NeurIPS 2023 · similarity 0.88)
- Cross-Modal Distillation for 2D/3D Multi-Object Discovery from 2D Motion (CVPR 2025 · similarity 0.87)
- SSLFusion: Scale and Space Aligned Latent Fusion Model for Multimodal 3D Object Detection (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
