TUM

[Sun-Fan]

hello, thanks for your paper. In your paper, you have used 8 Tums.I want to ask if this will cause the low fps. After all, there are many Convolution layers.Thanks!

[muye5]

Both TUM and SFAM are a little complex，it doesn't make sense why this architecture could speed up so fast. Both the hourglass TUM and fully connection attention are not friendly to speed. Wait for the code.

[qijiezhao]

TUM is a Thinned U-shape Module, it is not as slow as you think. The parameters of A TUM is even less than a Conv layer 1024x3x3x1024.
SFAM is also not as complex. First, concatenation; then implement SE attention for multi-scale features. Only 6 SE attentions. As a comparison, SE-ResNet101 have more and more SE attentions than ResNet101.

[leo-XUKANG]

i have tried it in darknet ,and where the network get deeper ,then the network will be slower,

[HoracceFeng]

@leo-XUKANG very interesting try. I am also curious why M2DET will have such fast speed. Would you share your code of M2DET in darknet?