Research map / Detection and segmentation / 3D detection and driving
CaMuViD: Calibration-Free Multi-View Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CaMuViD, a calibration-free multi-view pedestrian detector that learns feature projection and back-projection matrices with a fine-tuned InternImage backbone and cascade R-CNN head, detecting directly in image space.
Problem
Multi-view pedestrian detection in crowds relies on camera calibration and BEV ground-plane projection, which causes feature distortion, overfitting, and poor generalization across camera setups.
Approach
Replace geometric calibration-based projection with learned fully connected networks that estimate per-view projection matrices (and their inverses as back-projection matrices) on feature maps, fuse views in a common space by concatenation, and detect bounding boxes directly in each camera image.
Key results
- Achieves MODA of 95.0% on Wildtrack (vs. 94.1% for MVFP) and 96.5% on MultiviewX (tied with TrackTacular, 0.8 points above MVFP's 95.7%, the paper's stated margin).
- Cross-dataset (MultiviewX to Wildtrack) MODA of 86.4% vs. 76.7% for MVFP.
- Replacing learned projections with camera-parameter projections drops Wildtrack MODA by 9.1%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CaMuViD instead of MVFP | Multi-view pedestrian detection · Wildtrack | MODA: 94.1 -> 95.0 | Helped |
| Use CaMuViD instead of MVFP | Multi-view pedestrian detection · MultiviewX | MODA: 95.7 -> 96.5 | Helped |
| Train on MultiviewX and test on Wildtrack with CaMuViD instead of MVFP | Cross-dataset multi-view detection · Wildtrack (trained on MultiviewX) | MODA: 76.7 -> 86.4 | Helped |
| Replace summation fusion with concatenation (ablation) | CaMuViD without FRM · Wildtrack | MODA: 93.8 -> 94.6 | Helped |
| Add the Feature Refinement Module after concatenation fusion (ablation) | CaMuViD · Wildtrack | MODA: 94.6 -> 95.0 | Helped |
| Add the Feature Refinement Module after concatenation fusion (ablation) | CaMuViD · Wildtrack | MODP: 80.2% to 80.9% | Helped |
| Replace learned projection matrices with camera parameter-based projections (ablation) | CaMuViD · Wildtrack | MODA: 95.0% -> 9.1% | Hurt |
| Remove camera views, using only C1 instead of all seven cameras (ablation) | CaMuViD · Wildtrack | MODA: 95.0 -> 60.1 | Hurt |
| Use only cameras C1–C2 instead of all seven cameras (ablation) | CaMuViD · Wildtrack | MODA: 95.0 -> 77.8 | Hurt |
| Use only cameras C1–C3 instead of all seven cameras (ablation) | CaMuViD · Wildtrack | MODA: 95.0 -> 90.6 | Hurt |
| Use only cameras C1–C4 instead of all seven cameras (ablation) | CaMuViD · Wildtrack | MODA: 95.0 -> 93.8 | Hurt |
| Use only cameras C1–C5 instead of all seven cameras (ablation) | CaMuViD · Wildtrack | MODA: 95.0 -> 93.8 | Hurt |
Surprising findings
- Cross-dataset MODP drops to 60.7%, lower than most baselines, despite higher MODA and recall.
Limitations
- Detection-only; no multi-view association or tracking yet.
- MODP is IoU-based, making it not directly comparable with BEV Euclidean MODP and lower under cross-dataset shift.
Experimental setup
Datasets: Wildtrack MultiviewX
Benchmarks: Wildtrack MultiviewX
Metrics: MODA MODP precision recall F1
Baselines: MVDet Deep-Occlusion DeepMCD POM-CNN SHOT MVDeTr MVAug 3DROM MVFP TrackTacular GMVD
Builds on: InternImage MVDet Cascade R-CNN
Closest papers by meaning
- Towards Generalizable Multi-Camera 3D Object Detection via Perspective Rendering (AAAI 2025 · similarity 0.89)
- Unified Domain Generalization and Adaptation for Multi-View 3D Object Detection (NeurIPS 2024 · similarity 0.86)
- CorrBEV: Multi-View 3D Object Detection by Correlation Learning with Multi-modal Prototypes (CVPR 2025 · similarity 0.85)
- GeoBEV: Learning Geometric BEV Representation for Multi-view 3D Object Detection (AAAI 2025 · similarity 0.84)
- SparseAlign: a Fully Sparse Framework for Cooperative Object Detection (CVPR 2025 · similarity 0.84)
- Multi-View Attentive Contextualization for Multi-View 3D Object Detection (CVPR 2024 · similarity 0.84)
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.