FCHD-Fully-Convolutional-Head-Detector

Code for FCHD - A fast and accurate head detector

This is the code for FCHD - A Fast and accurate head detector. See the paper for details and video for demo.

Dependencies

The code is tested on Ubuntu 16.04.
install PyTorch >=0.4 with GPU (code are GPU-only), refer to official website
install cupy, you can install via pip install cupy-cuda80 or(cupy-cuda90,cupy-cuda91, etc).
install visdom for visualization, refer to their github page

Installation

Install Pytorch
Clone this repository

git clone https://github.com/aditya-vora/FCHD-Fully-Convolutional-Head-Detector

Build cython code for speed:

cd src/nms/
python build.py build_ext --inplace

Training

Download the caffe pre-trained VGG16 from the following link. Store this pre-trained model in data/pretrained_model folder.
Download the BRAINWASH dataset from the official website. Unzip it and store the dataset in the data/ folder.
Make appropriate settings in src/config.py file regarding the updated paths.
Start visdom server for visualization:

python -m visdom.server

Run the following command to train the model: python train.py .

Demo

Download the best performing model from the following link.
Store the head detection model in checkpoints/ folder.
Run the following python command from the root folder.

python head_detection_demo.py --img_path <test_image_name> --model_path <model_path>

Results

Method	AP
Overfeat - AlexNet [1]	0.62
ReInspect, Lfix [1]	0.60
ReInspect, Lfirstk [1]	0.63
ReInspect, Lhungarian [1]	0.78
Ours	0.70

Runtime

Runs at 5fps on NVidia Quadro M1000M GPU with 512 CUDA cores.

Acknowledgement

This work builds on many of the excellent works:

Reference

[1] Stewart, Russell, Mykhaylo Andriluka, and Andrew Y. Ng. "End-to-end people detection in crowded scenes." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.

ckhui/FCHD-Fully-Convolutional-Head-Detector