Project V3.1 — Crowd Density Estimation with CSRNet

Real-time crowd density heatmap and count estimation using CSRNet
(VGG-16 frontend + dilated CNN backend) on ShanghaiTech Part A & B.

---

Demo

---

Results

Dataset	Crowd Type	MAE ↓	RMSE ↓	Test Images
ShanghaiTech Part A	Dense	87.85	137.21	182
ShanghaiTech Part B	Sparse	21.12	37.97	316

Trained from scratch on M1 MacBook Air CPU. No GPU used.

---

Architecture

Input Image
    │
    ▼
Frontend: VGG-16 (first 10 conv layers, pretrained ImageNet)
    │         ↓ stride 8
    ▼
Backend: 6 × Dilated Conv (dilation=2) → 1×1 Conv
    │
    ▼
Density Map (H/8 × W/8) — sum = predicted crowd count

---

What I Learned

• Density map regression as an alternative to direct object counting
• Adaptive vs fixed Gaussian kernel generation for ground truth maps
• Dilated convolutions for multi-scale context without resolution loss
• MAE/MSE evaluation for count estimation tasks
• Real-time density heatmap overlay pipeline on video streams

---

Dataset

ShanghaiTech — Part A (dense, 482 images) and Part B (sparse, 716 images)
Download: Google Drive

After downloading, place at:

datasets/ShanghaiTech/
├── part_A/
│   ├── train_data/{images, ground_truth}
│   └── test_data/{images, ground_truth}
└── part_B/
    ├── train_data/{images, ground_truth}
    └── test_data/{images, ground_truth}

---

Setup

conda create -n csrnet python=3.10 -y
conda activate csrnet
pip install torch torchvision torchaudio
pip install opencv-python scipy matplotlib PyYAML tqdm

---

Usage

1 — Generate Density Maps

python scripts/generate_density_maps.py --config configs/config.yaml --part both

2 — Train

# Train Part B first (faster), then Part A overnight
caffeinate -i python src/train.py --config configs/config.yaml --part B
caffeinate -i python src/train.py --config configs/config.yaml --part A

3 — Evaluate

python src/evaluate.py --config configs/config.yaml

4 — Video Demo

# Place crowd video at inputs/crowd_video.mp4, then:
python scripts/inference_video.py --config configs/config.yaml --part A
# Output: outputs/video/demo_output.mp4

---

Tech Stack

Component	Choice
Framework	PyTorch
Frontend	VGG-16 (torchvision, pretrained)
Backend	Dilated CNN (6 layers, dilation=2)
Dataset	ShanghaiTech Part A & B
Kernels	Adaptive (Part A) / Fixed (Part B)
Training	M1 CPU, batch=1, Adam, 200 epochs
Inference	M1 MPS
Environment	conda, Python 3.10

---

Project Structure

crowd-density-estimation-csrnet/
├── configs/config.yaml         # all parameters — no hardcoded values
├── src/
│   ├── dataset.py              # CrowdDataset with augmentation
│   ├── model.py                # CSRNet architecture
│   ├── train.py                # training loop
│   ├── evaluate.py             # MAE/MSE + visualisations
│   └── utils.py                # shared utilities
├── scripts/
│   ├── generate_density_maps.py
│   ├── inspect_dataset.py
│   ├── validate_density_maps.py
│   ├── check_model.py
│   ├── inference_video.py
│   └── plot_training_curves.py
└── outputs/
    ├── evaluation/             # figures, CSVs, comparison panels
    └── video/                  # demo MP4