物体检测中的训练样本采样 直播笔记



  • 物体检测中的训练样本采样 直播笔记

    嘉宾:商汤科技研究副总监 陈恺

    一、分析

    分类得到的score与目标得分不一定同高同低,比如:

    image-20190919201812482

    image-20190919202249149

    二、提出HLR

    image-20190919203003707

    当D的IoU从0.92 -> 0.81时,D的重要性应该是不变的。

    image-20190919203124509

    image-20190919203316792

    image-20190919211623872

    image-20190919203650877

    image-20190919204115598

    三、提出PISAimage-20190919204230157

    (1)ISRimage-20190919204425556

    image-20190919204459626

    (2)CARL

    考虑两个分支的耦合关系:image-20190919204604743

    添加分支$L_{carl}$来抑制不重要的sample造成的影响,其中f(${s_i}$)为predict confidence![image-20190919204640773](/Users/yang/Library/Application Support/typora-user-images/image-20190919204640773.png)

    image-20190919204906303

    (3)PISA的效果(效果好,更适配多尺度变化):

    image-20190919205336703

    左边为正样本的ISR,横坐标为前几的top框,可见达到了期望,重要的样本loss大。image-20190919205425324

    image-20190919205815750

    image-20190919205923882

    image-20190919205933142


 

Copyright © 2018 bbs.dian.org.cn All rights reserved.

Looks like your connection to Dian was lost, please wait while we try to reconnect.