Research map / Detection and segmentation / 3D detection and driving
Leveraging Vision-Centric Multi-Modal Expertise for 3D Object Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use VCD-A with the 256×256 BEV feature configuration | BEVDepth-based camera-only detector · nuScenes val | NDS: 0.542 -> 0.566 | Helped |
| Use VCD-A with ConvNeXt-B | BEVDepth-based camera-only detector · nuScenes test | NDS: 0.610 -> 0.631 | Helped |
| Use VCD-E as the expert instead of a heterogeneous expert (ablation) | BEVDepth apprentice · nuScenes val expert comparison | mAP: 0.292 -> 0.354 | Helped |
| Replace BEVDistill with VCD-A distillation (ablation) | VCD-E to BEVDepth apprentice · nuScenes val distillation comparison | NDS: 0.439 -> 0.459 | Helped |
| Replace ResNet-50 with ConvNext-B (ablation) | VCD-E · nuScenes val backbone comparison | mAP: 0.611 -> 0.664 | Helped |
| Use LiDAR depth where available and predicted depth otherwise (ablation) | VCD-E · nuScenes val depth fusion | mAP: 0.495 -> 0.646 | Helped |
| Add the VCD distillation framework (ablation) | BEVDepth apprentice · nuScenes val, temporal length 8 | NDS (%): 40.9 -> 45.9 | Helped |
| Increase trajectory supervision from 5 to 9 points (ablation) | VCD-A · nuScenes val, temporal fusion length 8 | NDS (%): 45.9 -> 44.7 | Hurt |
| Add trajectory-based distillation (ablation) | VCD-A with temporal fusion and occupancy reconstruction · nuScenes val component ablation | NDS: 0.416 -> 0.449 | Helped |
| Add long-term temporal fusion (ablation) | VCD-A · nuScenes val component ablation | mAP: 0.271 -> 0.297 | Helped |
| Add occupancy reconstruction and long-term temporal fusion (ablation) | VCD-A · nuScenes val component ablation | mAP: 0.297 -> 0.308 | Helped |
| Distill full BEV features without trajectory-based supervision (ablation) | VCD-A · nuScenes val, trajectory length 0 | NDS (%): 40.9 -> 42.1 | Helped |
Closest papers by meaning
- SimDistill: Simulated Multi-Modal Distillation for BEV 3D Object Detection (AAAI 2024 · similarity 0.92)
- DI-V2X: Learning Domain-Invariant Representation for Vehicle-Infrastructure Collaborative 3D Object Detection (AAAI 2024 · similarity 0.88)
- PromptDet: A Lightweight 3D Object Detection Framework with LiDAR Prompts (AAAI 2025 · similarity 0.88)
- GAFusion: Adaptive Fusing LiDAR and Camera with Multiple Guidance for 3D Object Detection (CVPR 2024 · similarity 0.87)
- VeXKD: The Versatile Integration of Cross-Modal Fusion and Knowledge Distillation for 3D Perception (NeurIPS 2024 · similarity 0.87)
- Weak-to-Strong 3D Object Detection with X-Ray Distillation (CVPR 2024 · similarity 0.87)
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.