P2RLoss

Official code for CVPR-2025 paper:"Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting."

Overview

This repository implements the P2RLoss method for semi-supervised point-based crowd counting. The project uses a teacher-student framework with VGG16BN backbone and proposes a novel Point-to-Region Loss to improve counting performance with limited labeled data.

Features

• Semi-supervised learning: Utilizes both labeled and unlabeled data for training
• Teacher-student framework: Employs EMA-based teacher model for pseudo-labeling
• P2RLoss: Novel loss function that matches points to regions for better supervision
• Strong augmentation: Uses strong data augmentation for unlabeled samples
• Multi-stage training: Two-stage training strategy with progressive semi-supervised learning

Requirements

• Python 3.x
• PyTorch >= 1.8
• torchvision
• timm
• yacs
• numpy
• PIL
• opencv-python
• tqdm

Installation

# Clone the repository
git clone https://github.com/Elin24/P2RLoss.git
cd P2RLoss

# Install dependencies
pip install torch torchvision timm yacs numpy opencv-python tqdm

Data Preparation

The project uses ShanghaiTech dataset. Please organize your data as follows:

data/
└── ShanghaiTech/
    └── part_A/
        ├── train_data/
        │   ├── images/
        │   └── new-anno/
        └── test_data/
            ├── images/
            └── new-anno/

You also need to prepare a protocol file that specifies which images are labeled for semi-supervised learning. The protocol file should contain image IDs (one per line).

Training

Basic Usage

python main.py --data-path /path/to/data \
    --label 5 \
    --protocol /path/to/protocol.txt \
    --batch-size 16 \
    --tag experiment_name

Using the provided script

Modify run.sh with your data paths and configurations:

datadir=/path/to/ShanghaiTech/part_A
name=sha
part=5
T=${name}-L${part}

mkdir exp
mkdir exp/$T
mkdir exp/$T/code
cp -r datasets exp/$T/code/datasets
cp -r models exp/$T/code/models
cp -r losses exp/$T/code/losses
cp ./*.py exp/$T/code/
cp run.sh exp/$T/code

mkdir exp/$T/train.log
python main.py --data-path $datadir \
    --label ${part} --protocol /path/to/protocol.txt \
    --batch-size 16 --tag $T 2>&1 | tee exp/$T/train.log/running.log

Arguments

• --data-path: Path to the dataset
• --label: Percentage of labeled data (e.g., 5 for 5%)
• --protocol: Path to the protocol file
• --batch-size: Batch size for training
• --tag: Experiment tag for logging
• --resume: Path to checkpoint for resuming training
• --eval: Evaluation mode only

Configuration

The training configuration is managed through config.py. Key parameters include:

• DATA.BATCH_SIZE: Batch size (default: 1)
• DATA.DATASET: Dataset name (default: 'shha')
• MODEL.NAME: Model architecture (default: 'VGG16BN')
• MODEL.LOSS: Loss function (default: 'P2R')
• TRAIN.EPOCHS: Total training epochs (default: 1500)
• TRAIN.BASE_LR: Base learning rate (default: 5e-5)
• TRAIN.LR_SCHEDULER: Learning rate scheduler settings

Code Structure

P2RLoss/
├── config.py              # Configuration management
├── main.py                # Main training and evaluation script
├── run.sh                 # Training script
├── utils.py               # Utility functions
├── logger.py              # Logging utilities
├── lr_scheduler.py        # Learning rate scheduler
├── datasets/              # Dataset loaders
│   ├── __init__.py
│   ├── shha.py           # ShanghaiTech dataset
│   └── utils.py          # Dataset utilities
├── models/                # Model architectures
│   ├── __init__.py
│   ├── vgg16bn.py        # VGG16BN backbone
│   └── utils.py          # Model utilities
└── losses/                # Loss functions
    ├── __init__.py
    └── p2rloss.py        # Point-to-Region Loss

Model Architecture

The model uses VGG16BN as the backbone with:

• Multi-scale feature extraction
• Feature fusion layer
• Decoder for density map generation
• Two output channels for density estimation

Training Strategy

1. Stage 1 (Epochs 0-25): Supervised training with labeled data only
2. Stage 2 (Epochs 25-50): Semi-supervised training with both labeled and unlabeled data
   • Teacher model generates pseudo-labels for unlabeled data
   • Student model learns from both labeled and pseudo-labeled data
   • Progressive weight adjustment for semi-supervised loss

Results

The model achieves competitive performance on ShanghaiTech dataset with limited labeled data. Training logs and curves are saved to exp/<tag>/train.log/.

Citation

If you find this code useful for your research, please cite:

@inproceedings{lin2025point,
  title={Point-to-Region Loss for Semi-Supervised Point-Based Crowd Counting},
  author={Lin, Wei and Zhao, Chenyang and Chan, Antoni B},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={29363--29373},
  year={2025}
}