Closed martinmatak closed 5 years ago
I think that a smaller network could achieve reasonable result in estimating age groups. I'm not sure but the group 0-20 might be better to be broken down into finer groups because of high variance in appearance.
That said, would you assume that VGG16 that is pretrained on ImageNet should with fine tuning on appa test dataset (and this age grouping task) be able to have an accuracy of >0.5? (given that every group has equal number of samples)
On Sat, 23 Feb 2019 at 18:49, Yusuke Uchida notifications@github.com wrote:
I think that a smaller network could achieve reasonable result in estimating age groups. I'm not sure but the group 0-20 might be better to be broken down into finer groups because of high variance in appearance.
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/yu4u/age-gender-estimation/issues/83#issuecomment-466673071, or mute the thread https://github.com/notifications/unsubscribe-auth/AHUndg8xt5qrDSwQ5hZKz8g_VMlUtboeks5vQX81gaJpZM4bODp7 .
I did not do that; but it might be possible to achieve the accuracy.
https://github.com/yu4u/age-gender-estimation/issues/48#issuecomment-410283362
interesting, thanks! :)
Hi,
sorry for raising an issue! (I know it is not a code issue, but more a question. I didn't know how else to do it.)
I'd like to come up with a network that is smaller than complex ResNet-50/VGG16, yet able to estimate at least an age group (0-20, 21-40, 41-60, ... or even better 0-10, 11-20, ... or at least 0-30, 31-60, 61-100) of a 224x224x3 image.
Do you know maybe if there is any work on this? What do you think about this problem? Is this achievable?
Thanks for your answer!