Human Segmentation in Surveillance Video with Deep Learning

This repo contains the official training and testing code for Human Segmentation in Surveillance Video with Deep Learning (Multimedia Tools and Applications, 2020).

Please, refer to the paper for technical details.

A video demo of our project is available on the project page.

Usage

The implementation of our work is built on MATLAB R2018a and the Deep Learning Toolbox is requested. Note that this tutorial assumes that your root folder is /human-segmentation/.

Please modify the commands where appropriate if you choose to use different directories. Download the dataset and the source code. Your file structure should look like this:

/human-segmentation/
    dataset/
        imageDataset/
            train/
            test/
            val/
        pixelLabelDataset/
            train/
            test/
            val/
    myColorMap.m
    pixelLabelColorbar.m
    preprocessImage.m
    semanticseg_newImage.m
    testMySegnet.m
    trainMySegnet.m

Train

Run the code in trainMySegnet.m to train the network.

Test

Run the code in testMySegnet.m to test the trained network and compute metrics.

Pretrained models

• Download the pretrained models. Please, note that they can only be used for non-commercial uses.
• Choose the suitable network: trainedNetwork_indoor.mat and trainedNetwork_outdoor.mat are the trained networks for indoor and outdoor areas that we chose.
• Load the appropriate model into the MATLAB command window and then query the network with images in the "photos" folder, which contains real photos taken directly with a camera.

Query trained network

Run the code in semanticseg_newImage.m to query the trained network and get the network output and the label overlay, which is an overlap of input and output based on a colormap. The colormap is defined in myColorMap.m.

The function pixelLabelColorbar.m is useful to add a colorbar to the current axis. The colorbar is formatted to display the class names with the colors.

The function preprocessImage.m is used to resize the input image before passing it to the network, according to the image size defined in the input layer.

Dataset

We will release our dataset for encouraging future research on human segmentation. Please send an email to monica.gruosso@unibas.it if you need it for academic research and non-commercial purposes.

Before requesting our data, please verify that you understand and agree to comply with the following:

• This data may ONLY be used for non-commercial uses (This also means that it cannot be used to train models for commercial use).
• You may NOT redistribute the dataset. This includes posting it on a website or sending it to others.
• You may include images from our dataset in academic papers.
• Any publications utilizing this dataset have to reference our paper.
• These restrictions include not just the images in their current form but any images created from these images (i.e. “derivative” images).
• Models trained using our data may only be distributed (posted on the internet or given to others) under the condition that the model can only be used for non-commercial uses.

Citation

If you use the code or the data for your research, please cite the paper:

@article{gruosso2020human,
  title={Human segmentation in surveillance video with deep learning},
  author={Gruosso, Monica and Capece, Nicola and Erra, Ugo},
  journal={Multimedia Tools and Applications},
  pages={1--25},
  year={2020},
  publisher={Springer}
}