MDnet visual tracking algorithm implementation version 3. A trainded model mdnetv3.pt which is trained on vot2016 and one result, vot2016/marching are also uploaded, in which blue windows are groundtruth boundingboxes while green ones are tracking results. Better performance with respect to fps, precision and success has achieved
How to run:
python srcv3.py online0
the program asks you to input a video name and you need to download and prepare vot2016 datasets
Files:
libv3.py: contains all classes and most functions
options.py: contains all parameters we need to modify
srcv3.py: offline_training, online_tracking
Folder paths organization:
-vot2016/
-mdnet/
-libv3.py
-srcv3.py
-options.py
-vot2016.txt
-results/
-trained_nets/
-mdnetv3.pt
Dependencies:
(1)python3.5
(2)opencv,numpy
(3)pytorch
(4)scikit-learn
Hardware:
Nvidia GTX TITAN X(recommended), but it can also run without gpu. Offline training needs 1390 M graphics memory and online tracking needs about 8000 M graphics memory