基于深度学习的图像匹配技术专题- [Patch-Based Matching 1]-从matchnet 开始

MatchNet:Unifying Feature and Metric Learning for Patch-Based Matching【CVPR15, github】

数据下载：论文章中使用的数据库的链接变了，想下载数据库的朋友可以点击我是patchdata.

Firstly, the model is built based on caffe frame. U can obtain the paper and source code on the Internet.  The architecture provided in the paper is divided into two parts：A (Feature network) and B (metric network), which can be regard as extracting and descriptor.

If you want to train this net by yourself, you can finish it though following steps.

##cd data/phototour   and   download liberty.zip  

unzip -q -d liberty liberty.zip 

rm liberty.zi

##Generate leveldb database for liberty and solve it in "data/leveldb"

./run_gen_data.sh  #脚本调用genetate_patch_db.py

DATASET=liberty
python generate_patch_db.py data/phototour/${DATASET}/info.
data/phototour/${DATASET}/interest
data/phototour/${DATASET} data/leveldb/${DATASET}.leveldb

A. Feature network

B. Metric network