search

msracver / Deformable-ConvNets

Deformable Convolutional Networks
MIT License
4.05k stars 957 forks source link

How to get the receptive field? #282

Closed kirkzZ closed 3 years ago

kirkzZ commented 3 years ago

I want to get like this image

jsjxyjl commented 3 years ago

I don't know. Was it photoshopped

HaozhiQi commented 3 years ago

For example, the coordinate of the top blue dot is (x, y). Then you could get its sampling location (\delta x, \delta y) from the offset feature map. The result (x + \delta x, y + \delta y) is plotted in the middle layer. Because it's 3x3 kernel, there are 9 points plotted in the middle layer.

The process is done again for the bottom (9 x 9) points.

HaozhiQi commented 3 years ago

Also see #10 and https://github.com/msracver/Deformable-ConvNets/blob/master/rfcn/deform_conv_demo.py

kirkzZ commented 3 years ago

Also see #10 and https://github.com/msracver/Deformable-ConvNets/blob/master/rfcn/deform_conv_demo.py

The version is too old, is there a new version to implement