Open xonobo opened 6 years ago
Yeah, it is intentional, because author removes 38x38 feature map, so use 19x19 feature map twice.
It is an intentional design. But if you want to improve the performance of the small object detection, my suggestion is to use 38x38 without ResBlock instead of using 19x19 twice. The actual speed is almost the same.
How do I pass 38x38 feature map instead of the duplicate 19x19 feature map?
According to dot file I obtained for the deploy_merged.prototxt the 19x19 feature maps are fed to confidence and location extraction layers twice. Is it intentional or a copy/paste error?
deploy_merged.pdf