GOTURN Training Toolkit
This is the code for training of GOTURN tracker implemented inside OpenCV.
Original GOTURN paper:
Learning to Track at 100 FPS with Deep Regression Networks,
David Held, Sebastian Thrun, Silvio Savarese,
European Conference on Computer Vision (ECCV), 2016 (In press)
Install Caffe and compile using the CMake build instructions: http://caffe.berkeleyvision.org/installation.html
sudo apt-get install libopencv-dev
Building HDF5 dataset
For GOTURN training, first HDF5 dataset shold be generated. It can be done by simple function call:
By default it generates 10 HDF5 datasets every with 500x10 samples (10 crop samples per image as proposed by authors). As alternative more low-level function can be used:
GOTURN training requires a "bvlc_reference_caffenet.caffemodel" for GOTURN network weights inialization:
Training is launched by next line:
All hyperparameters are configured in goturnSolver.prototxt, for more details refer Caffe documentation.
Evaluate the tracker
In order to visualize results, GOTURN tracker can be tested on test dataset:
buildH5Datasets("D:/ALOV300++/trainDataset.h5", 10); testNet("goturn_iter_30000.caffemodel");
First command generate a new small test dataset, and second launching a test procedure with visualization.
Pretrained GOTURN model
Also there is pretrained GOTURN model is available in OpenCV_extra repository