CodePudding user response:
Although the SAMF and DSST can keep up with the normal target dimension change, but only seven SAMF scale is coarse, and there are 33 DSST scale comparative fine accuracy; DSST test best translation first, then detect the best scale, is the best step by step, and SAMF is translational scale test, translation and scaling is optimal at the same time, but often local optimum and global optimum is not the same; DSST will track is divided into translation tracking and scale with two problems, can use different methods and characteristics, more flexible, but need additional training a filter, each frame size detection need 33 sampling image block, after characteristics are calculated respectively, add window, such as FFT, scale filter a much slower than the translational filter; SAMF, need only a filter does not require additional training and storage, each scale detection on a feature and FFT, but in the image block is larger than the amount of calculation DSST is high, so the scale detection DSST is not always better than the SAMF, actually in VOT2015 and VOT2016 SAMF are more than DSST, of course, this is mainly because characteristic is better, but at least scale method is not poor, on the whole, DSST is very novel, faster and SAMF also good is also more accurate,