地标检索和识别 比赛介绍分析: https://blog.csdn.net/sunflower_sara/article/details/99361693
地标检索 方法A :https://blog.csdn.net/sunflower_sara/article/details/100549946
文件保存情况 :https://blog.csdn.net/sunflower_sara/article/details/100549920
其他方案 https://blog.csdn.net/sunflower_sara/article/details/100549909
涉及到的内容:
地标检索:利用多个backbone(inception)网络和GeM Pooling+ Triplet Loss提取全局网络特征描述子;建立KDTree通过KNN寻找近邻,利用近邻根据权重重新生成新的描述子,以进行数据增强提高召回率,建立新的KDTree,将近邻返回为初步检索结果;最后, 根据深度局部特征(DELF)建立新的tree,找到近邻点,并由随机抽样一致(RANSAC)算法进行几何验证(geometric verification),去除误匹配点,选择局内点数多的为目标,对检索结果进行重排。
GeM Pooling
https://blog.csdn.net/sunflower_sara/article/details/100555479
Triplet Loss
https://blog.csdn.net/jcjx0315/article/details/77160273
https://blog.csdn.net/weixin_40671425/article/details/98068190
&Focal loss
https://blog.csdn.net/u014380165/article/details/77019084
https://zhuanlan.zhihu.com/p/49981234
https://zhuanlan.zhihu.com/p/32423092
https://www.cnblogs.com/king-lps/p/9497836.html
https://blog.csdn.net/Code_Mart/article/details/89736187
CE LOSS https://blog.csdn.net/sunflower_sara/article/details/81162964
深度学习常用损失函数 https://blog.csdn.net/Tianlock/article/details/88232467
KNN 和KdTree
https://blog.csdn.net/eeeee123456/article/details/79927128
https://blog.csdn.net/csdn0123zl/article/details/81253648
https://blog.csdn.net/luoyexuge/article/details/84235421
https://blog.csdn.net/ye1215172385/article/details/80214776
https://www.jianshu.com/p/4fd7cad27906
Kmeans
https://blog.csdn.net/weixin_37536446/article/details/81326932
DELF
https://blog.csdn.net/wangxinsheng0901/article/details/81906696
RANSAC
https://blog.csdn.net/robinhjwy/article/details/79174914
四个步骤:
1.Global CNN descriptors:
对于imageNet的tensorflow-slim Inceptionv3进行finetune
GeM pooling
batch hard Triplet Loss 三个部分的loss
2.Approximate Nearest Neighborbor Search 近邻搜索
Ann开源包
3.Database augmentation (Query Expansion):
权重衰减和10个近邻 去代替图像的描述子 logarithmic weights
weights = logspace(0,-1.5,10)
4.Re-ranking based Delf
DeLF(Deep Local Features 深度局部特征)
DELF :
GemPooling
介于mean pooling和max pooling之间,二者是其特殊形式
通过调节参数p,可以关注不同细度的区域
公式: