5fa1a76
1
2
The authors use the features generated after passing these regions through a pre-trained CNN like ResNet as visual embeddings.
