search

chenhsuanlin / inverse-compositional-STN

Inverse Compositional Spatial Transformer Networks :performing_arts: (CVPR 2017 oral)
MIT License
318 stars 63 forks source link

IC-STN algorithm for landmark detection #14

Closed SunnyDp closed 5 years ago

SunnyDp commented 5 years ago

Thanks for your great job. I want to use the IC-STN for the face landmark detection task. Different from the classification task, do I need to inverse transform the transformed image to obtain the points on the orginal image？Waiting for your advice.

chenhsuanlin commented 5 years ago

I'm not exactly sure what you mean, could you please elaborate?

SunnyDp commented 5 years ago

@chenhsuanlin Sorry for the unclear question. In the classification task such as MNIST, the transformed images are obtained to get more accurate result, but for regression task, such as face landmark detection, the aim is to get the point coordinates on the face. So if I use the IC-STN to transform the image and detect the points on the transformed image, the coordinates are not on the original image. How can I deal with it? Is it necessary to inversely transform the coordinates?

wuneng commented 5 years ago

@SunnyDp i have the same problem with you

chenhsuanlin commented 5 years ago

Yes, that sounds right to me -- as long as your warp function is invertible (e.g. affine or homography), you can simply invert the parameters from IC-STN and apply them on the facial landmarks. (cc'ing @wuneng also)