Dear Organisers,
If an image contains more than one faces with different labels (gender, smile), how do you evaluate. Besides, there are two end times in the “phase” (March 15 and March 30), which one is true?
Dear kpzhang,
Thank you for you post.
Regarding your questions:
1. All images should have only one face. They have all been cropped so there is only one face in each image.
2. As for the date, thank you for bringing in to our attention. The final date is the 30th of March. We have now changed it.
Let us know if you have any further questions. Good luck!
Kind Regards,
MTT
Posted by: MTT @ March 4, 2016, 9:16 a.m.Thanks for your reply. I am still confused about the testing images. Is there large background in any images? I think an image with complex background and small face is more regular. If no, we need not to do face detection and localization, but I think you should provide the style of cropping, because it is important to many algorithms.
Could you provide a testing example (cropped images), it is the best way to understand the rule? Thanks!
Bests,
Kaipeng
Hi Kaipeng,
The test set will be generated in the exact same way as the training and validation set (same tools, same sources for the images, same cropping techniques, same sizes, same distribution, same backgrounds, etc). At this stage we can't disclose any samples, though. However, you can expect them to be identical to the training and validation sets that we have made public.
Let me know if you have any further questions.
Kind Regards,
MTT
Posted by: MTT @ March 4, 2016, 12:56 p.m.Thanks for your reply. I find that there are more than one face in many images (many faces is unlabeled) both trainning and validation. So, I want to know how do you crop images to ensure that all testing images have only one face.
Posted by: kpzhang @ March 4, 2016, 1:19 p.m.For example, there are two faces in smiles_valset/im_GenFex_08540.jpg. What should I output.
Posted by: kpzhang @ March 4, 2016, 1:31 p.m.Hi Kaipeng,
In the cases where you might find more than one face in the images, you should use the bounding box provided in the csv file. The coordinates in that file will let you know which face is the one you should be using in your training.
I hope this makes sense. Do let me know if you have any further questions.
Thank you and good luck!
Kind Regards,
MTT
Posted by: MTT @ March 8, 2016, 1:06 p.m.Thanks for your reply. We don’t know which testing image have more than one face. So, can our code output all candidate faces with their labels? And you select one face we provided according to the annotated bounding box. Thanks!
Bests,
Kaipeng
More specific, can our code output all candidate faces’ bounding boxes (coordinates) with their labels (smile and gender) in an image?
Posted by: kpzhang @ March 8, 2016, 2:27 p.m.Hi Kaipeng,
You should only give us one prediction per image. If your face locator returns more than one face, choose the one which is closer to the centre of the photo. All of the training, validation and testing images will be centred around the face whose attributes we are interested in.
Let me know if you have further questions.
Thank you.
MTT
Posted by: MTT @ March 9, 2016, 11:41 a.m.