FgSegNet : Foreground Segmentation Network

This repository contains source codes and training sets for the following paper:

"Foreground Segmentation Using Convolutional Neural Networks for Multiscale Feature Encoding" by Long Ang LIM and Hacer YALIM KELES

The preprint version of the above paper is available at: https://arxiv.org/abs/1801.02225

Citation

If you find FgSegNet useful in your research, please consider citing:

@article{LIM2018256,
        title = "Foreground segmentation using convolutional neural networks for multiscale feature encoding",
        journal = "Pattern Recognition Letters",
        volume = "112",
        pages = "256 - 262",
        year = "2018",
        issn = "0167-8655",
        doi = "https://doi.org/10.1016/j.patrec.2018.08.002",
        url = "http://www.sciencedirect.com/science/article/pii/S0167865518303702",
        author = "Long Ang Lim and Hacer Yalim Keles",
        keywords = "Foreground segmentation, Background subtraction, Deep learning, Convolutional neural networks, Video surveillance, Pixel classification"
}

Requirements

This work was implemented with the following frameworks:

• Spyder 3.2.x (recommended)
• Python 3.6.3
• Keras 2.0.6
• Tensorflow-gpu 1.1.0

Usage

Easy to train! Just a single click, go!

1. Clone this repo: git clone https://github.com/lim-anggun/FgSegNet.git

2. Modify the following file:

   • <Your PYTHON 3.6>\site-packages\skimage\transform\pyramids.py
     in pyramid_reduce function, replace the following two lines
     out_rows = math.ceil(rows / float(downscale))
out_cols = math.ceil(cols / float(downscale))
     with
     out_rows = math.floor(rows / float(downscale))
out_cols = math.floor(cols / float(downscale))

3. Download VGG16 weights from Here and put it in FgSegNet/FgSegNet/ dir, or it will be downloaded and stored in /.keras/models/ automatically.

4. Download CDnet2014 dataset, then put it in the following directory structure:
   

   Example:

     FgSegNet/
          FgSegNet/FgSegNet_M_S_CDnet.py
                  /FgSegNet_M_S_SBI.py
                  /FgSegNet_M_S_UCSD.py
                  /FgSegNet_M_S_module.py
          SBI2015_dataset/
          SBI2015_train/
          UCSD_dataset/
          UCSD_train20/
          UCSD_train50/
          FgSegNet_dataset2014/
                    baseline/
                         highway50
                         highway200
                         pedestrians50
                         pedestrians200
                         ...
                     badWeather/
                         skating50
                         skating200
                         ...
                    ...
           CDnet2014_dataset/
                     baseline/
                          highway
                          pedestrians
                          ...
                     badWeather/
                          skating
                          ...
                     ...
   

5. There are two methods; i.e. FgSegNet_M and FgSegNet_S. Choose a method that you want to train by setting method_name=='FgSegNet_M' or method_name=='FgSegNet_S'.

6. Run the codes with Spyder IDE. Note that all trained models will be automatically saved (in current working directory) for you.

Evaluation

on CDnet2014 dataset

We perform two separated evaluations and report our results on two test splits (test dev & test challenge):

1. We compute our results locally. (on test dev dataset)
2. We upload our results to Change Detection 2014 Challenge. (on test challenge dataset where ground truth values are not shared with the public dataset)
   (Both results are reported in our paper. Please refer to it for details)
   

Compute metrics locally using changedetection.net > UTILITIES tab.

Note:

• test dev: by considering only the range of the frames that contain the ground truth labels by excluding training frames (50 or 200 frames)
• test challenge: dataset on the server side (http://changedetection.net)

on SBI2015 dataset

We split 20% for training (denoted by n frames, where n ∈ [2−148]) and 80% for testing.

on UCSD Background Subtraction dataset

We perform two sets of experiment: first, we split the frames 20% for training (denoted by n frames, where n ∈ [3 − 23]) and 80% for testing, second we split 50% for training (where n ∈ [7 − 56]) and remaining 50% for testing.

Results

Results on CDnet2014 dataset

Table below shows overall results across 11 categories obtained from Change Detection 2014 Challenge.

Methods	PWC	F-Measure	Speed (320x240, batch-size=1) on NVIDIA GTX 970 GPU
FgSegNet_M	0.0559	0.9770	18fps
FgSegNet_S	0.0461	0.9804	21fps

Results on SBI2015 dataset

Table below shows overall test results across 14 video sequences.

Methods	PWC	F-Measure
FgSegNet_M	0.9431	0.9794
FgSegNet_S	0.8524	0.9831

Results on UCSD Background Subtraction dataset

Tables below show overall test results across 18 video sequences.

For 20% split

Methods	PWC(th=0.4)	F-Measure(th=0.4)	PWC(th=0.7)	F-Measure(th=0.7)
FgSegNet_M	0.6260	0.8948	0.6381	0.8912
FgSegNet_S	0.7052	0.8822	0.6273	0.8905

For 50% split

Methods	PWC(th=0.4)	F-Measure(th=0.4)	PWC(th=0.7)	F-Measure(th=0.7)
FgSegNet_M	0.4637	0.9203	0.4878	0.9151
FgSegNet_S	0.5024	0.9139	0.4676	0.9149

Updates

07/08/2018:

• combine FgSegNet_S with FgSegNet_M and more
• add SBI2015, UCSD dataset and more results
• add published paper and a new citation
• integrate custom loss and other codes in a single file, for ease of use
• other improvements

09/06/2018:

• add quantity results on changedetection.net (CDnet2014 dataset)
• add supporting codes

29/04/2018:

• add Jupyter notebook and a model for testing (FgSegNet_M).

27/01/2018:

• add FgSegNet_M (a triplet network) source code and training frames

Contact

lim.longang at gmail.com
Any issues/discussions are welcome.