Closed uestcMeng closed 3 years ago
Hi @uestcMeng, many thanks for your interest in RT-GENE!
Re MPII: I honestly can't remember. From their dataset description, it looks like they provided the eye area with corresponding eye landmark positions for both eyes. From that we probably extracted eye patches for both left and right eyes.
Re augmentation: Please refer to the code, which should be easy to understand. Let me know if you have further questions after checking the code.
Best wishes, Tobias
thanks you, Tobias! Your reply helps me a lot!
Hi, I notice that you use MPII dataset in your paper, and employ the both eye patches as input. I find that MPII dataset contains 1500 images of the left and right eye respectively, but the left and right eye are not paired. And I also find there is another dataset called MPIIFace, which provide full face image, so I wonder whether you use MPIIFace rather than MPII. Besides, you proposed a method of image augmentation, "we perform 10 augmentations by cropping the image on the sides and subsequently resizing it back to its original size. Each side is cropped by a pixel value drawn independently from a uniform distribution U(0, 5)." I have some doubts here,do you mean to randomly crop the eye image ten times and then resize it back to origin size? Would you mind answering my questions?Thanks a lot!