Time: January 26th, 2018, the nineteenth week
Weekly Summary: In this week, I learn how to use the GPU server to run the code such as ECO algorithm. Futhemore, I continued reading a paper about discriminative tracking. This paper is an almelioration of SRDCF. This is also from the same author Martin Danelljan.
Paper: Adaptive Decontamination of the Training Set: A Unified Formulation for discriminative Visual Tracking/ Author: Martin Danelljan, Gustav Hager, Fahad Shahbaz Khan, Michael Felsberg/ Publication information: CVPR 2016
Outline: This algorithm is the improvement of SRDCF. They try to up-weight the correct samples and down-weight the corrupted ones during tracking classification. Because the bad samples is contaminated by background clutter, occlusion or motion blur. This unified formulation is based on the traditional discriminative classifier. But it gives weight to every samples in one frame and different weight to previous frames. Those samples value is continuous and normalized. When adding the two parameters in the regressor, it becomes a biconvex optimization problem. Consequently, they use Alternate Convex Search to fix one parameter and optimize another. And this optimization could be implemented by the function in matlab.
Advantages: This method is unified and could be employed in other classifier to solve the drift problem. And it performs better than other decontamination methods in their experiments. The experiments is very sufficient to prove the algorithm's good performance. The algorithm does improve robustness and accuracy.
Disadvantages: They discard the previous frames in this work. It may cause over-fitting problem. And this iteration procedure in this algorithm is time consuming.