search

yangxue0827 / R2CNN_FPN_Tensorflow

R2CNN: Rotational Region CNN Based on FPN (Tensorflow)
419 stars 139 forks source link

Does the constant [103.939, 116.779, 123.68] have any effection on training? #45

Closed eccstartup closed 6 years ago

eccstartup commented 6 years ago

I read about this constant in your paper that it is the mean value of training set, and the reference1.py runs pretty well with sample pictures. But is it the general setting for any set of dataset? I mean, how to set this value for my own dataset, or, I can just ignore it?

I kind of have trouble in training. It just won't converge. With limitted number of pictures(144 for training actually), how to adjust parameters and superparameters of this training?

Thanks.

engineer1109 commented 6 years ago

No effection

eccstartup commented 6 years ago

@engineer1109 How did you know? Have you tried train1.py?

engineer1109 commented 6 years ago

@eccstartup [103.939, 116.779, 123.68] means the pixel mean value. No use.

yangxue0827 commented 6 years ago

[103.939, 116.779, 123.68] is from ImageNet, you don not need to change it when you train your own dataset. What's more, this trick does not effect a lot.