Structure about Feature map in SSD-Mobilenet

[liekejiang]

I found that the first feature layer output is 1919 and last two layers both give a 11 size output, dose it correct? Why choosing such a small size comparing to the original structure?

[lzx1413]

It follows the structure in mobilenet parper.

[liekejiang]

@lzx1413 I wonder that if get the feature layer from the previous layer in mobilenet, than we could get larger feature layer, maybe we could get better result

[lzx1413]

@liekejiang Despite the computation cost, features with feature stride 8 may have not large enough receptive field to handle the smallest object in the image. Besides, features with stride 8 are so shallow that may not have enough semantic presentations. But this will also enlarge the search space for objects, which is good for the recall rate. You can try this to see the result at last.