Jie Zhang

level-2 and level-3 use random shift around the ground truth. rotation and flip can also be applied.

The model I trained may use iters 1000,000 or 2000,000. The loss will still reduce when training goes on. Maybe you should plot the loss to see if it is....

I wonder if your detector can find out this kind of faces. I'm not sure whether the network can predict landmarks or not, since the training data may be lack...

What if level-1 is failed, the data processed by level-1 are not stable, there might be some wrong data for level-2. Randomly sample around a point can cover the situation...

maybe you should run more iterations or tuning the learning rate and other parameters during the training goes on.

Because of the network initialization, there's no guarantee that the loss will be stable or able to reach a point gives you a low loss. You should consider to restart...

more data can help training the network but the network initialization part has nothing to do with your data. Actually, the default weight initialization method is ok for most situation...

@wuqiangch the result here is 3 level ? and the level-1 includes EN and NM ? what's more, did you monitor the loss during network training? Maybe some cnns are...

the training data you have put out is ok. 5 landmark has 10 values and image data is a multi-dimension variable so it may be hard to view all values....

the model I provided under webapp/v0.0.6, did it have the same problem?