菜鸟实习日记~day23(NUIST 2016)

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

4.同时也考虑了上下文的重要性