Скачать с ютуб Implementation of MDNet, KCF and SiamFC в хорошем качестве

Implementation of MDNet, KCF and SiamFC

This video demonstrates the implementation of three recent visual object trackers: KCF, SiamFC and MDNet. The green box is the ground-truth bounding box while the red box is the estimated target position according to the tracker. I modified KCFpy (https://github.com/jbhewitt12/KCFpy) and py-MDNet (https://github.com/jbhewitt12/py-MDNet) to output accuracy and robustness according to the VOT Challenge standards (https://arxiv.org/abs/1503.01313) I modified a SiamFC to track objects with a webcam, and also added an automatic face recognition feature using the cv2 library. Analysis: ** Matrix sequence ** MDNet Frame Rate: 1.92 KCF Frame Rate: 197 KCF is particularly susceptible to failure in the presence of visual noise. This is demonstrated in the Matrix sequence, where KCF scores an accuracy of 40% and requires 3 reinitializations. MDNet scores an accuracy of 53% and only requires 1 reinitialization. The major breakthrough of KCF is that it can analyse thousands of negative samples per frame while standard trackers sample only a few. This is a significant advantage in most cases where the background contains a lot of varied information, but if the background samples are all similar due to heavy noise then this advantage is negated. *Skating sequence* MDNet Frame Rate: 1.80 KCF Frame Rate: 175 This was one of the few sequences in which the Accuracy of KCF was higher than MDNet. This is an example of where the focus of KCF on massive negative sampling shines; sequences with a varied high-information background and where the positive samples of the target may not be reliable as it undergoes significant appearance changes. *SiamFC object and face tracking* In these sequences I attempted to demonstrate the robustness of SiamFC by including occlusion, rotation, appearance changes and scale changes of the targets in the two videos. Music: Kozoro - Breathe   / breathe