Compactness Driven Co-learning for Crowd Counting and Localization

Official PyTorch implementation of "Compactness Driven Co-learning for Crowd Counting and Localization".

Installation

1. Create Conda Environment

conda create -n cclm python=3.10 cmake=3.16.3
conda activate cclm

2. Install Dependencies

pip install -r scripts/requirements.txt

Dataset Preparation

Please follow FIDTM to process the datasets into COCO format.

The expected directory structure is:

data/
├── part_A_final/
│   ├── train_data/
│   ├── test_data/
│   ├── train_data_annotation.json
│   └── test_data_annotation.json
├── part_B_final/
├── NWPU/
└── FDST/

Training

1. Modify Configuration

Edit configs/fidt_ucl.json:

• Change Dataset.train.name to the target dataset: sta_crowd, stb_crowd, nwpu_crowd, or fdst_crowd
• Update ann_file and img_prefix to your actual dataset paths

2. Launch Training

For single-node multi-GPU training (4 GPUs):

export CUDA_VISIBLE_DEVICES=0,1,2,3
torchrun --nproc_per_node=4 --nnodes=1 --master_port 12350 train_counter.py --config configs/fidt_ucl.json

Evaluation

1. Compile NMS Module

Before evaluation, compile the Cython NMS module:

python setup.py build_ext --inplace

2. Run Evaluation

export CUDA_VISIBLE_DEVICES=0
python inference_nwpu.py --config configs/fidt_ucl.json --ckpt path/to/your/best.pth

Replace path/to/your/best.pth with the actual path to your trained checkpoint.