search

luoyetx / deep-landmark

Predict facial landmarks with Deep CNNs powered by Caffe.
BSD 3-Clause "New" or "Revised" License
282 stars 156 forks source link

Hard to repeat level3's mean error at some points #14

Open mzpmzk opened 8 years ago

mzpmzk commented 8 years ago

Hi, thanks for your code sharing! I run your code several times by using bootstrap.sh. The result is very close to your presentation in level1 and level2, but the mean error in level3 is rising randomly at some points. (eg: Nose-0.121428, Left Mouth-0.062793 level3) Could you give me some suggestions to solve this problem? Thanks a lot！

luoyetx commented 8 years ago

@mzpmzk Maybe you need to carefully train the networks of level-3. Actually, I train the networks not by bootstrap.sh but using caffe train for every networks. You need to pay attention to caffe's output, mainly the loss, if it's not stable, the network may not train well, In this case, you should retrain the network.

mzpmzk commented 8 years ago

@luoyetx Thank you very much for your reply! I will try.

pidanxhq commented 8 years ago

@luoyetx After testing the train results of level2 and level3,the accuracy for right eye is worse,do you konw why ? (others are all ok)And I have check the train network.It is right.waiting for your reply! level3:mean error :Right Eye = 0.061876 failure:Right Eye = 0.909117 level2:mean error :Right Eye = 0.015816 failure:Right Eye = 0.031737

luoyetx commented 8 years ago

@pidanxhq pay attention to the training log, your networks for Right Eye doesn't not converge.

pidanxhq commented 8 years ago

level3_test This it shows.I do not know how it happens

pidanxhq commented 8 years ago

@luoyetx What does the red line mean.