How is the pretrained VGG loaded?

[oblivious137]

The pre-trained VGG contains parameters of the two FCs after pooling. I want to know whether they are loaded before oicr training?

If so, how is this VGG pre-trained? The pretrained model seems to contain a ROIPooling layer, while the original VGG contains a GlobalPooling. Hope to point out which paper it comes from.

[oblivious137]

Sorry, after checking the VGG paper, I found that the input of FC is 51277, 7=224/2^5, which is determined by the input size 224 and the number of maxpools 5. So, are these two FCs loaded to OICR model ?

[ch13d2am]

I'm not the developer, but if you check the vgg_weights_helper module in the utils folder, you will seethat all the weights for the vgg layers are loaded. But in the vgg16.py file under models, the FCs are not included - this is exactly as it is in the OICR paper. Only the conv layers are used as feature extractor, and then passed with the selective search proposals into first SPP and then FRCNN for detection.