MPSN

Code for MPSN: Motion-aware Pseudo Siamese Network for Indoor Video Head Detection

Paper

Dependencies

• The code is tested on Ubuntu 20.04.2,python 3.8,cuda 10.1.

• install torch version

  pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 torchaudio==0.8.0 -f https://download.pytorch.org/whl/torch_stable.html

• install cupy, you can install via

  pip install cupy-cuda101

Installation

1. Install pytorch

2. Clone this repository

git clone https://github.com/pl-share/MPSN

3. Install

pip install -r requirements.txt

Datasets

• Download the restaurant dataset from the following link. Unzip it and store the dataset in the data/ folder, You can also change the data set loading path by modifying src/config.py.

• RGBdata_flow.zip : The original and optical flow images

• RGBdata_diff.zip : The original and frame-difference images

Eval

• Download model from the following link.
• Store the head detection model in checkpoints/output/ folder.
• if you want to eval other model,you should modify head_backbone.py simultaneously

        addnet = right_res()  #line 59
        left_vgg = left_res() #line 60
        in_channel = 256      #mob:96 res:256

• Run the following python command from the root folder.

python eval.py --model_path <model_path>

​ Restaurant datasets model

model	test ap	Anchor scale	pertrained model
flow mob DFA	0.790	[2,4]	True
diff mob DFA=+	0.838	[2,4]	True
diff resnet DFA	0.802	[2,4]	True
diff vgg DFA	0.824	[8,16]	True
diff vgg DFA=+	0.857	[8,16]	True

• If you load 'diff vgg DFA+APC', please modify eval.py

from src.head_detector_vgg16 import Head_Detector_VGG16 #line 14
from trainer import Head_Detector_Trainer   #line 15

head_detector_mpsn = Head_Detector_VGG16(ratios=[1], anchor_scales=[8, 16])    #line 142					
#head_detector_mpsn = mob(ratios=[1], anchor_scales=[2,3])     #line 143

• If you choose DFA=+ +APC , please modify train_or.py and head_detector.py simultaneously

#hf2 = t.mul(h1, t.sigmoid(h2)) + h2
hf2 = h1+h2

• If backbone is vggnet, please modify trainer.py and head_detector1.py simultaneously

Training

python train.py

Citation

Please refer to the following bibtex to cite.

@article{SUN2022109354,
         title = {MPSN: Motion-aware Pseudo-Siamese Network for indoor video head detection in buildings},
         journal = {Building and Environment},
         volume = {222},
         pages = {109354},
         year = {2022},
         issn = {0360-1323},
         doi = {https://doi.org/10.1016/j.buildenv.2022.109354},
         author = {Kailai Sun and Xiaoteng Ma and Peng Liu and Qianchuan Zhao}
}

Contact Us

If you have other questions❓, please contact us in time 👬

Acknowledgement

This work builds on the excellent work:

• FCHD-Fully-Convolutional-Head-Detector by Aditya Vora