search

WongKinYiu / yolov7

Implementation of paper - YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors
GNU General Public License v3.0
13.22k stars 4.18k forks source link

SPP vs SPPCSP #817

Open iumyx2612 opened 1 year ago

iumyx2612 commented 1 year ago

SPPCSP is more computation heavy than normal SPP, is it better in term of AP?

yulin010101 commented 1 year ago

You can refer to https://arxiv.org/abs/2011.08036.

iumyx2612 commented 1 year ago

You can refer to https://arxiv.org/abs/2011.08036.

This is the neck structure of YOLOv7 image And this is the SPPCSPC module image if I replace the SPPCSPC with SPP, it should be much faster since the SPPCSPC has two extra 3x3 Conv.
Also, the 2 pathway has the same channels as the input feature maps (c=1024), according to normal CSP-ized, it should be half of the channel of the input feature maps, which is 512 no?

fn-hide commented 1 year ago

You can refer to https://arxiv.org/abs/2011.08036.

This is the neck structure of YOLOv7 image And this is the SPPCSPC module image if I replace the SPPCSPC with SPP, it should be much faster since the SPPCSPC has two extra 3x3 Conv. Also, the 2 pathway has the same channels as the input feature maps (c=1024), according to normal CSP-ized, it should be half of the channel of the input feature maps, which is 512 no?

How can we make the visualization of this architecture? I mean, i confuse about the line. Where should i connect the layers.

PascAlex commented 1 year ago

How did you manage to understand the structure of the neck ? I am reading the paper and still can't understand Even backbone and head