Research map / 3D vision / 3D generation
Multi-View People Detection in Large Scenes via Supervised View-Wise Contribution Weighting
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 a supervised view-wise contribution weighting network for multi-view people detection in large scenes, trained on the synthetic CVCS dataset and adapted to new scenes with 5% labels plus a simple discriminator-based domain adaptation.
Problem
Multi-view people detection methods are trained and evaluated on small single scenes with few frames and fixed cameras, so they generalize poorly to larger, more complex scenes with occlusion and calibration error.
Approach
Each view's predicted ground-plane people map (supervised by the view's field-of-view ground truth derived from scene GT) is passed through a small subnet to produce per-view fusion weight maps, which are normalized over visible views and used for weighted feature fusion, requiring no extra labels.
Key results
- On CVCS, the method achieves MODA 46.2 and F1 68.4, the best F1 among compared methods (SHOT F1 67.0).
- On CityStreet, F1 of 76.0 is second only to 3DROM (79.2); the method has the best average rank (1.5) across CVCS and CityStreet.
- On Wildtrack, cross-scene finetuned model with domain adaptation reaches MODA 78.9 and F1 89.5, below single-scene SOTA 3DROM (F1 96.7).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use supervised view-wise contribution weighting rather than the MVDet fusion approach | Ours on CVCS · CVCS | F1_score: 60.9 -> 68.4 | Helped |
| use supervised view-wise contribution weighting rather than the MVDet fusion approach | Ours on CityStreet · CityStreet | F1_score: 68.4 -> 76.0 | Helped |
| add supervised view-wise contribution weighted fusion (ablation) | Ours with ResNet · CVCS | F1_score: 60.9 -> 68.4 | Helped |
| add supervised view-wise contribution weighted fusion (ablation) | Ours with VGG · CVCS | F1_score: 59.4 -> 61.5 | Helped |
| use supervised rather than unsupervised view-wise fusion (ablation) | Ours · CVCS | F1_score: 66.6 -> 68.4 | Helped |
| use supervised rather than unsupervised view-wise fusion (ablation) | Ours · CityStreet | F1_score: 73.1 -> 76.0 | Helped |
| test with a variable number of camera views after training on 5 views (ablation) | Ours · CVCS | F1_score: 66.1, 68.4, 69.7, 68.8 | Mixed |
| add domain adaptation after finetuning on 5% of target-scene training images (ablation) | Ours (ft) · Wildtrack | F1_score: 87.0 -> 89.5 | Helped |
| add domain adaptation after finetuning on 5% of target-scene training images (ablation) | Ours (ft) · MultiviewX | F1_score: 90.1 -> 91.4 | Helped |
Surprising findings
- The cross-scene method with only 5% target labels and domain adaptation remains below single-scene state-of-the-art methods trained on 100% of the same scene's data on Wildtrack and MultiviewX.
- 3DROM, a data augmentation method, performs much worse than the proposed method on CVCS but best on CityStreet.
Limitations
- The domain adaptation method is simple and limited by image style transfer.
- Cross-scene results on Wildtrack and MultiviewX trail single-scene state-of-the-art methods.
- Relies on finetuning with 5% of labeled target-scene data.
Experimental setup
Datasets: CVCS CityStreet Wildtrack MultiviewX
Benchmarks: CVCS CityStreet Wildtrack MultiviewX
Metrics: MODA MODP Precision Recall F1_score Avg. Rank
Baselines: MVDet SHOT MVDeTr 3DROM RCNN POM-CNN DeepMCD DeepOcc Volumetric
Builds on: MVDet SHOT MVDeTr 3DROM
Closest papers by meaning
- WILDTRACK: A Multi-Camera HD Dataset for Dense Unscripted Pedestrian Detection (CVPR 2018 · similarity 0.85)
- Multi-view Crowd Tracking Transformer with View-Ground Interactions Under Large Real-World Scenes (CVPR 2026 · similarity 0.84)
- Learning from Synchronization: Self-Supervised Uncalibrated Multi-View Person Association in Challenging Scenes (CVPR 2025 · similarity 0.84)
- MTMMC: A Large-Scale Real-World Multi-Modal Camera Tracking Benchmark (CVPR 2024 · similarity 0.82)
- CaMuViD: Calibration-Free Multi-View Detection (CVPR 2025 · similarity 0.82)
- Unified Domain Generalization and Adaptation for Multi-View 3D Object Detection (NeurIPS 2024 · similarity 0.82)
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.