Optimized MDNet for visual object tracking

This repository contains a MatConvNet re-implementation for MDNet algorithm, which is ~10x and ~6x faster than the original matlab and python implementations, respectively.

• MDNet-matlab-Org

• MDNet-python

Detail Comparisons

                |-------------------------------------------------------------------|
                |           | MDNet | pyMDNet | MDNet-Org (Ours) | MDNet-Opt (Ours) |
                |-------------------------------------------------------------------|
                | OTB-2015  |  67.9 |  65.2   |       66.4       |      67.2        |
                |-------------------------------------------------------------------|
                | VOT-2015  |  37.8 |   --    |       36.8       |      39.3        |
                |-------------------------------------------------------------------|
                | FPS (OTB) |   ~1  |   ~2    |       ~13        |      ~13         |
                |-------------------------------------------------------------------|

• MDNet: the original matlab implementation

• pyMDNet: python implementation

• MDNet-Org (Ours): our implementation using default parameters (see setting_mdnet_org)

• MDNet-Opt (Ours): our implementation using our settings (see setting_mdnet_opt)

All trackers are benchmarked on OTB-2015 dataset using a single GPU (GTX 1080).

Requirements and Dependencies

• NVIDIA GPU with compute capability 3.5+
• Matlab 2017a or above
• MatConvNet

Quick Start

To run pre-trained MDNet for OTB testing, please follow these steps:

1. Clone this repository into $MDNet:

   git clone git@github.com:ZjjConan/Optimized-MDNet.git $MDNet

2. Complie your MatConvNet

3. Change paths

• setup_optmdnet:

  lib_path for your own matconvnet

• run_evaluation_OPE:

  savePath for your tracking results

  videoPath for OTB dataset

  videoAttr for OTB subset (OTB2013 or OTB2015)

4. Models

   mdnet_vot_otb: training on VOT13/14/15 datasets for OTB testing.

   mdnet_otb_vot: training on OTB dataset for VOT15 testing.

5. For VOT testing

   • copy files in vot/vot_tracker_settings into your own vot workspace.

   • changes paths in tracker_OptMDNet_Opt or tracker_OptMDNet_Opt for

     tracking with different parameters.

Training Your Own Model

please find detailed settings in pretraining fold for database setup and network training.

Citations

If you use this project in your research, please cite the original MDNet paper:

@InProceedings{nam2016mdnet,
    author = {Nam, Hyeonseob and Han, Bohyung},
    title = {Learning Multi-Domain Convolutional Neural Networks for Visual Tracking},
    booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    month = {June},
    year = {2016}
}

License

This software is being made available for research purpose only. Check LICENSE file for details.