Research map / Detection and segmentation / Open-set and few-shot detection
Embodied Crowd Counting
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.
Introduces Embodied Crowd Counting (ECC), a UE4-based simulator dataset (ECCD) of large outdoor crowd scenes, and ZECC, a zero-shot drone agent that uses MLLM-driven altitude switching and normal-line-based viewpoint selection to count crowds.
Problem
Crowd counting from passive cameras cannot resolve occlusion in large-scale scenes, and existing embodied navigation benchmarks are small indoor or lack large crowd object quantities.
Approach
Treat crowd counting as an active embodied task: a zero-shot agent uses an MLLM (GPT-4V) to decide between high-altitude and low-altitude exploration, then computes top-down viewpoints along surface normals of crowd clusters to reduce occlusion while keeping travel cost low.
Key results
- ZECC achieves 18.71% MAPE with 3722.45 m travel distance on ECCD, versus 46.01% MAPE for the best ZSON baseline (CoW + GD), though with higher TD than FBE.
- Compared to multi-view counting, ZECC reaches 18.71% MAPE using 5 cameras, while MVF-10 and CountFormer-10 reach 15.13% and 12.8% with 1735 cameras.
- In real-world tests on two outdoor scenes, NLBN + GD reduces MAPE to 15.17% and 19.83% from 73.74% and 78.44% without NLBN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ZECC instead of CoW + GD | ECCD crowd counting · ECCD | MAPE (%): 46.01 ± 0.96 -> 18.71 ± 1.41 | Helped |
| use ZECC instead of FBE + GL | ECCD exploration · ECCD | TD (m): 2513.06 ± 247.35 -> 3722.45 ± 73.78 | Hurt |
| use ZECC instead of CountFormer-10 | ECCD multi-view counting comparison · ECCD | MAPE (%): 12.8 ± 0.00 -> 18.71 ± 1.41 | Hurt |
| use ZECC instead of CountFormer-10 | ECCD multi-view counting comparison · ECCD | # of Cameras: 1735.32 ± 0.00 -> 5 ± 0.00 | Helped |
| add high-altitude exploration (ablation) | ZECC ATE · ATE ablation | MAPE (%): 17.46 -> 18.91 | Hurt |
| add high-altitude exploration (ablation) | ZECC ATE · ATE ablation | TD (m): 4633.67 -> 3804.63 | Helped |
| add low-altitude exploration (ablation) | ZECC ATE · ATE ablation | MAPE (%): 88.08 -> 18.91 | Helped |
| add NLBN (ablation) | ZECC crowd counting · NLBN ablation | MAPE (%): 65.19 -> 18.91 | Helped |
| add normal-line navigation (ablation) | ZECC NLBN · NLBN ablation | MAPE (%): 98.44 -> 18.91 | Helped |
| add viewpoint selection (ablation) | ZECC NLBN · NLBN ablation | MAPE (%): 92.55 -> 18.91 | Helped |
| add ATE-view-vector-based viewpoint selection (ablation) | ZECC NLBN · NLBN ablation | Successful rate (%): 1.45 -> 100.00 | Helped |
| apply NLBN (ablation) | GL crowd counting · Real environment, Scenario 1 | MAPE (%): 41.41 -> 22.50 | Helped |
Surprising findings
- Crowd counting models (GL) generally underperform detection models (GD) for ZECC as agents get closer to sparse crowds, attributed to training on dense images.
Limitations
- MLLM planning fails when building occlusion hides the crowd during high-altitude exploration.
- Only static crowds considered; dynamic targets and real-world deployment left for future work.
- Performance depends on hyperparameters (cluster size, navigation vector degree, density threshold) and the simulator may have biases.
Experimental setup
Datasets: ECCD (simulated, Unreal Engine 4)
Benchmarks: ECCD
Metrics: MAPE travel distance (TD) successful rate
Baselines: FBE + GL FBE + GD CoW + GL CoW + GD OpenFMNav + GL OpenFMNav + GD MVF CountFormer
Builds on: Zero-shot object-goal navigation Frontier-based exploration Generalized Loss crowd counting Grounding DINO
Closest papers by meaning
- Cross-View Cross-Scene Multi-View Crowd Counting (CVPR 2021 · similarity 0.86)
- Adapting Lightweight Image-based Counting Models for Video Crowd Counting (CVPR 2026 · similarity 0.85)
- Counting without Scale: Scale-Consistent Error Correction for Crowd Counting (NeurIPS 2026 · similarity 0.84)
- Learning From Synthetic Data for Crowd Counting in the Wild (CVPR 2019 · similarity 0.84)
- DecideNet: Counting Varying Density Crowds Through Attention Guided Detection and Density Estimation (CVPR 2018 · similarity 0.84)
- Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting (CVPR 2025 · similarity 0.83)
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.