Abstract
This paper proposed a scalable logo detection method and a very large logo dataset called WebLogo-2M. The aim of the author is realistic application
Method
To establish the large logo dataset, the author scratch the images with logo key words in Twitter. The final dataset WebLogo-2M includes 194 classes and totally more 1800000 images.
To address the problem that lots of noisies in web images, they use three screening measures.
About deep network, they use incremental learning to repeatedly self-train.
My Opinion
The manner of solving the imbalance of different classes is inspired.
