文章方法:
1.Classification Model:Google net(smaller and faster) 相比较于VGG
2.使用Cascade Region Regression

(1)Increase the variation of initialization during training.
(2)Augment training data by removing 0.7 NMS step around GT after RPN.
3.Tracking:使用相关滤波

(1)Increase the variation of initialization during training.
(2)Augment training data by removing 0.7 NMS step around GT after RPN.