《Learning Deep Features for Scene Recognition using Places Database》记录
***scene dataset:
scene15: 15 categories, a few hundred images per class, 95% accuracy.
MIT indoor67: 67 categories, indoor scene
SUM db: 397 categories, 100 images per class
places db: 700w, 476 categories.
生成方式:
1. 生成section,来源于sum db。
2. amt,进行2轮label,(1) default no, choose yes. (2) from labeled yes image, default yes choose no. 90% accuracy. 750 image + 30 positive samples + 30 negative samples
***evalue db
a good database should be dense (with a high degree of data concentration), and diverse (it should include a high variability of appearances and viewpoints)