likelyzhao/NetVlad-MxNet

NetVlad layer with MxNet

NetVlad-MxNet

NetVlad Training with MxNet

steps:

1. prepare video scripts
2. cut frame and generate feature files
3. generate the label files
4. training

prepare video

json list. each line contains one json the format should be contain sevral key words like:

{
  "url":string,
  "type":"video",
  "label":
  {
    "name"：“video_terror”，
    “type“: "video",
    "version":"1",
    "data":[
            {
              "label":string,
              "segment":[float,float]# starttime,endtime
            },
            ...
     ]
  }
}

字段详细解释：

字段	类型	详细解释
url	string	视频的url路径（本地或者web）
type	string	类型 视频为video
name	string	标示，BK视频分类设置为：video_terror
version	string	版本信息，目前为1
label	string	BK的类别
segment	[float float]	开始和结束时间

cut frame and generate feature files @baobao

generate the label files  @baobao

training

netvlad symbol

dataiter

training hyper param