Density Adaptive Real-Time Crowd People Counting Network (DA-CSRNet)

Density-adaptive crowd counting for CCTV-like scenes: CSRNet + Self-Attention + Count Calibration (α)

---

Overview

This project implements a density-adaptive real-time crowd counting approach that improves robustness when the crowd density distribution in test scenes differs from training scenes.

While CSRNet achieves strong counting performance with a simple architecture (without multi-scale branches), it can suffer performance degradation under highly congested or density-shifted conditions.
To address this, we propose a system that:

1. augments CSRNet with Self-Attention to better capture global context and occlusion patterns, and
2. introduces an additional count calibration network that predicts a multiplicative factor α(X; φ) to adjust the final count.

---

Key Idea (Intuition)

Crowd counting models often output a density map and derive the final count by summing it. However, when density changes drastically (e.g., sparse ↔ highly congested), the same model may systematically over/under-estimate.

We treat this as a density-adaptation problem and learn an additional lightweight network that looks at image features and outputs α(X) to rescale the count:

• baseline count: C
• calibrated count: α(X) · C

---

Method Summary

1) CSRNet Backbone (Density Map Prediction)

Given an input image X_i, CSRNet predicts a density map:

• predicted density map: Z(X_i; θ)
• ground-truth density map: Z_i^{GT}

Training objective for CSRNet is the density-map regression loss:

$$ L_1(\theta) = \frac{1}{2N}\sum_{i=1}^{N} \left| Z(X_i;\theta) - Z_i^{GT} \right|_{2}^{2} $$

---

2) Self-Attention Module (Context Modeling)

We add Self-Attention so that the model can attend to:

• surrounding patterns,
• global layout,
• occlusion relationships in dense scenes.

This helps detect heads/people patterns more reliably when targets are heavily overlapped.

---

3) Count Calibration Network (Density-Adaptive Scaling)

We introduce a CNN-based calibration network that outputs:

• scaling factor: α(X_i; φ)
• calibrated count: α(X_i; φ) · C_i

The calibration network is trained using:

$$ L_2(\varphi) = \frac{1}{N}\sum_{i=1}^{N} \left| \alpha(X_i;\varphi)\cdot C_i - C_i^{GT} \right| $$

Where the CSRNet-derived count C_i is computed by summing the predicted density map:

$$ C_i = \sum_{l=1}^{L}\sum_{w=1}^{W} Z_{l,w} $$

So the final predicted crowd count becomes:

• Final Count: α(X_i; φ) · C_i

---

Evaluation

Dataset

• Shanghai dataset (as described in the paper)
• Total people instances: 330,165
• Total labeled images: 1,198

Metric: MAE

Baseline CSRNet MAE:

$$ MAE = \frac{1}{N_{test}}\sum_{i=1}^{N} \left| C_i - C_i^{GT} \right| $$

Proposed model MAE (using calibrated count):

$$ MAE = \frac{1}{N_{test}}\sum_{i=1}^{N} \left| \alpha(X_i;\varphi)C_i - C_i^{GT} \right| $$

---

Reported Result (from the paper)

On ShanghaiTech Part A test set:

• CSRNet: MAE 68.42
• Proposed (DA-CSRNet): MAE 67.98
• Improvement: −0.44 MAE

Note: This README reports the numbers as stated in the paper; reproduction may vary depending on implementation details, preprocessing, and training setup.

---

Roadmap

• Add training pipeline (CSRNet + Self-Attention)
• Add calibration network training (α)
• Add evaluation scripts (MAE, density map visualization)
• Provide pretrained weights & reproducibility notes

---

Citation

If you use this work, please cite the original paper (fill with your final bib entry):

@inproceedings{kim2023densityadaptive,
  title     = {Density Adaptive Real-Time Crowd People Counting Network},
  author    = {Kim, Jun-Hee and Kim, Chae-Ho and Kim, Dae-Seok and Lee, Suk-Ho},
  booktitle = {Proceedings of the Korea Multimedia Society (Autumn Conference)},
  year      = {2023}
}