(1) data, the data acquisition:
1. Benchmark is an industry benchmark (databases, articles, source, result).
2. excellent paper, usually experimental data set will introduce it uses, public data sets can be downloaded. When the application data sets, it is best to use the school's mailbox.
3. forums or exchange community, such as thinkface
4. The scale of the bigger the better data 2W
Binary data, a first face type, the second non-faces. Face Data: Path /ox.ipg_ 60,80,280,320-
non-face data: As long as the human face are possible.
For positive samples: Operation cropped. According to the coordinates of the marked face cut out. This tool can opencvy to complete the production of face data. To check, with a look at the data there is no problem.
For negative samples: for a random cutting, the IOU the ratio (overlap ratio) of the raw data in which, when the IOU <0.3, it is considered a negative sample, it is preferable to take data no human face as negative samples
produced LMDB data source (in fact, he is the Caffe supports very commonly used classification of data sources)
exzample and dATA folder settings are the same
tools is a caffe that where the tool of price stability bonds
train and val refers to the training and validation data sets location.
Production LMDB data source, caffe script file provided,
the first few lines into its own installation directory
conduct resize operation
size alenet and VGG are usually 227 * 227
