GOTURN-Tensorflow

This is a tensorflow implementation of GOTURN.

Thanks to author David Held for his help of this implementation.

The original paper is:

Learning to Track at 100 FPS with Deep Regression Networks,
David Held, Sebastian Thrun, Silvio Savarese,

The github repo for caffe implementation is given by the authors: davheld/GOTURN

Brief illustration of how this network works:

You can refer to the paper or github repo above for more details.

Environment

• python3
• tensorflow 1.0+, both cpu and gpu work fine

How to use it

Finetune for your own dataset

1. Create a folder, fill in all training images
2. Create a .txt file
   • It should contains target image, searching image and ground-truth bounding box
   • Bounding box is in the form of
     , usually from 0 to 1, but exceeding this range is also fine.
   • Example of one line:

     train/target/000024.jpg,train/searching/000024.jpg,0.29269230769230764,0.22233115468409587,0.7991794871794871,0.7608061002178649

3. Change related places in

   train.py

4. Train it and wait!

   python
   python train.py
   

5. The log file is

   train.log

   by default

Test

1. Download pretrained model from: GOTURN_MODEL
2. Uncompress the

   checkpoints

   folder, and put it in the root directory of this repo
3. Test on examples just by running

   load_and_test.py

   python
   python load_and_test.py
   

4. The log file is

   test.log

   by default

TIPS

Be careful, the output of this network actually always from 0 to 10 thus I multiplied the ground-truth bounding boxes( always ranging from 0 to 1) by 10.