about some details about m2det HOT 2 CLOSED

Good questions!
Recently, we're preparing for some dues, sorry for the late answering.
1, Why SFAM False: Two points: (1)SFAM will slow down the training process. (2)SFAM is a little tricky, sometimes positive while sometimes negative for the accuracy. For example, SFAM is beneficial for 320x320 size, but has no obvious improvement for 512x512 size. (we guess the main reason is training epochs). So, to avoid confusion, we set SFAM as False default.
2, the lack of initiation of Norm layer has no effect to the final accuracy.(You can add it) 40x40 is the largest feature, we follow SSD to norm it due to that normal distribution is beneficial to stabilize the training process.
3, Why set the fifth Conv layer (stride:1x1, padding:0x0): To leave a larger feature map for the minimum one. This step can improve the localization accuracy of large objects.
4, Is there any negative influence when resizing original images to square shape for object detection: Very little, because we employ the same operation at training process.
5, Where is multi-scale evaluation: I am busy for other dues recently, this script will appear in next month.
6, "the largest output feature map" of "The FFMv2 fuses the base feature and the largest output feature map of the previous Tum" in original paper is 40x40? Yes

from m2det.

@qijiezhao
many many thanks!!!

from m2det.