The ninth paper: C-COT--Beyond Correlation Filters: Learning Continuous Convolution Operators for Visual Tracking/ Author: Martin Danelljan, Andreas Robinson, Fahad Shahbaz Khan, Michael Felsberg / Publication information: ECCV 2016
Outline: This paper presents a novel framework to integrate the three layers of VGGNet in multi-resolution and utilize bicubic interpolation to obtain continuous convolution filters, labels and confidence maps. These continuous properties make the accurate location possible and robust. Firstly , they introduce the whole problem in to a T period problem in spatial domain, which is a hilbert space equipped with an inner product. When the interpolation operation is added into the loss function, they deploy a spatial regularization term to alleviate the boundary effect and a weight vector to strengthen the target samples. Further, they resort to Conjugate Gradient method to obtain the close-form solution. And they utilize grid search to get a routh solution and then utilize Newton; method to obtain the final location. Due to the favorable design in continuous spatial domain, the tracker is also suitable for feature point tracking, even sub-pixel location.
Methodology: 1. Features: CNN features from VGG Net. Layers 0,1,5. And multi-resolution of each layer.
2. The interpolation operator into continuous domain.
3. A spatial regularization term
4. Conjugate Gradient method.
5. Grid search and Newton' method.
Advantages: The best performance in that year. A very favorable fusion of multi layers in multi resolution. Down-weight the background to make the target models not be corrupted.
Disadvantages: Features and the relative filters in multi resolution is really time consuming. Maybe a multi-peak detection could promote its accuracy.