Mask_rcnn: What is the meaning of BACKBONE_STRIDES?

Created on 8 Jan 2018  路  3Comments  路  Source: matterport/Mask_RCNN

Is BACKBONE_STRIDES the feature map size in each stage?
If I change the image size from 128*128 to 512*512, should I change the value of BACKBONE_STRIDES?
Thanks!

Most helpful comment

Resnet50 has the same strides as Resnet101

All 3 comments

The backbone stride is the stride of each layer of the backbone network relative to the input image. For example, the C2 layer of the pyramid has a stride of 4, meaning that it's dimensions are 1/4 of the dimensions of the input image (i.e. if the input image is 1024x1024 the C2 size is 256x256).

Changing the input image size doesn't affect the backbone stride.

@waleedka : What is BACKBONE_STRIDES value when I use resnet50 model, instead of 101? Because your document said that

The strides of each layer of the FPN Pyramid. These values
are based on a Resnet101 backbone.
BACKBONE_STRIDES = [4, 8, 16, 32, 64]

Resnet50 has the same strides as Resnet101

Was this page helpful?
0 / 5 - 0 ratings