Keypoint head implementation

[LukasStu]

Hi Superlee,
thank you for your great work. I have a question about your keypoint mask head (build_fpn_keypoint_mask_graph).
As far as I can tell there are differences between yours the keypoint mask head discribed in the mask r-cnn paper. For example you used KEYPOINT_MASK_POOL_SIZE= 7 instead of KEYPOINT_MASK_POOL_SIZE= 14. You also used the deconv layer to reduce the feature map depth to num_keypoints.

Is there a certain reason for these changes?

In the picture you can see the head structure after those 8 conv layers.

Thanks,
Lukas