MFT
Multi-solution Fusion for Visual Tracking(MFT) The paper is coming.
Introduction
MFT tracker is based on correlation filtering algorithm. Firstly, we combine different multi-resolution features with continuous convolution operator~\cite{danelljan2017eco}. Secondly, we train multi-solution independently using different features and fuse multi-solutions optimally to predict target location, which drastically improves robustness. Respectively. At last, we reasonably choose different combinations of Res50, SE-Res50, Hog, CN features as our final feature to adapt to different tracking situation.
Usage
- Supported OS: the source code was tested on 64-bit CentOS Linux release 7.3.1611 (Core), and it should also be executable in other linux distributions.
- Dependencies:
- A modified version of matconvnet (included in the ./external_libs/matconvnet folder).
- autonn(https://github.com/vlfeat/autonn) is included in the (./external_libs/autonn) folder.
- MATLAB 2016b, and all different version will change a lot.
- Cuda 8.0 enabled GPUs
Preparation
cd ./feature_extraction/networks
wget http://www.vlfeat.org/matconvnet/models/imagenet-resnet-50-dag.mat
wget http://www.robots.ox.ac.uk/\~albanie/models/se-nets/SE-ResNet-50-mcn.mat
Setting export CUDA_CACHE_MAXSIZE=8000000000" in the ./~bash_profile so that gpuDevice(1) will take fewer time.
Demo
run demo_MFT.m()
VOT
[VOT Intergration] ./vot2018_main/MFT.m change ./tracker_MFT.m tracker_repo_path = 'your MFT path'