Receptive field calculation to get patch

songdejia / DFL-CNN

This is a pytorch re-implementation of Learning a Discriminative Filter Bank Within a CNN for Fine-Grained Recognition

MIT License

267 stars 52 forks source link

Receptive field calculation to get patch #16

Open pshroff04 opened 4 years ago

pshroff04 commented 4 years ago

Hi @songdejia ,

First of all great repo 👍 I am a little confused in this part of code. p_tl = (8*col, 8*row) p_br = (col*8+92, row*8+92)

I believe this comes from the paper "the output of the 10th convolutional layer conv4_3 represents patches as small as 92 × 92 with stride 8paper ". I didn't get how the author arrive here? Maybe a novice question but your help is appreciated. Thanks

ExNarc commented 4 years ago

It is a receptive field for VGG-16 at conv4_3. The size in conv4_3 is 56x56. and each px in 56x56 in represent the 92x92 px on origin input(448x448). you can take a look on this as a source: https://www.researchgate.net/figure/The-detailed-receptive-field-and-stride-size-in-original-VGG16-net_tbl1_333638990