Implementation of OverFeat for Object detection

[Nyan Naing]

Dear All,

I am trying to implement Object detection using OverFeat as discussed in thees two papers paper_1 and paper_2.

I have some points to be clear before implementation.

(1) How is feature response maps 20 x 15 x 4096 come out? Is it concatenation of all 355 x 355 pixels sub-images with 32 pixels strides and different scales?

I mean each 355 x 355 sub-image produces 1 x 1 x 4096 feature map and  20 x 15 x 4096 represents 20 x 15 = 300 sub-images of original 640 x 480 image?

(2) How is training image size? In the first paper, the context region size is bigger than detector size. So trained images are same as detector size?

(3)No proper discussion for bounding box regression.

Thanks