Point‑Based Crowd Counting (P2PNet)

Why this project

Counting methods based on density maps lose spatial precision in crowded scenes. This project implements a purely point‑based framework that directly predicts head locations under point supervision, emphasizing localization quality as much as counting accuracy.

Objective

Localize individuals and estimate counts with minimal supervision while remaining robust to extreme densities and perspective changes.

Method

• Direct point proposal network that outputs unordered sets of head centers.
• One‑to‑one bipartite (Hungarian) matching between predictions and ground truth to avoid duplicates and enforce precision.
• Loss combines classification and localization terms with a matching cost; training is stable without density maps or pseudo boxes.
• Density‑Normalized Average Precision (nAP) to evaluate localization under varying crowd densities.

Image ──► Backbone (e.g., ResNet) ──► Point Proposal Head ──► {p_i}
                                            │
                                            ├─ Matching (Hungarian) with GT points
                                            └─ Loss: cls + loc + λ·match

Data & training

• Supports standard crowd datasets; provides data loaders augmenting scale, perspective, and occlusion.
• Label‑efficiency sweeps (e.g., 100%, 50%, 10% points) to study supervision vs. performance.

Evaluation & ablations

• Metrics: MAE/MSE for counts, nAP for localization, and precision/recall at small radii.
• Ablations: matcher cost terms, NMS vs. set‑prediction, backbone depth, input resolution.

Outcomes

• Competitive MAE and strong nAP compared to density‑map baselines, particularly in high‑density regions.
• Evidence that point‑only supervision maintains localization fidelity and simplifies the pipeline.

Artifacts

• Clean training/eval scripts, nAP implementation, and visualizers for predicted point sets.