by HungWei-Andy @ NTU DISPLab

• Introduction
• Requirements
• Data Directory
• Initial VGG-M Model
• Pretrained Model
• Usage

Python (tensorflow) implementation of Hyeonseob Nam and Bohyung Han, Learning Multi-Domain Convolutional Neural Networks, CVPR2016.

Python 2.7
numpy>=1.12.1
tensorflow-gpu==1.0.0
matplotlib>=2.0.1
skimage>=0.13.0
Pillow>=2.2.1

Please download and put the video files of OTB and VOT in the directory 'data' with structure as follows:

.
├── data
|     ├── otb
|     |    ├── Basketball
|     |    └── ...
|     | 
|     └── vot
|          ├── vot2013
|          |      ├── bicycle
|          |      └── ...
|          | 
|          ├── vot2014
|          |      ├── ball
|          |      └── ...
|          |
|          └── vot2015
|                 ├── bag
|                 └── ...
|
├── models
└── README.md

The initial model is converted from caffe VGG-M model into .npy file using caffe-tensorflow library provided by ethereon.

To download the initial model, run download.sh directly.

  bash download.sh

We have pretrained the mdoel on vot dataset, otb dataset, and both. To download the pretrained model, run download_trained.sh directly.

  bash download_trained.sh

  python tracking.py --dataset dataset --seq sequence --load_path path [--no_display]

  bash pretrain.sh