my own data set
oneOfThePeople opened this issue · comments
hi,
i want to train my own data so i have some question.
- what is mean --crop-size and --origin-size like if my image in different size what is the origin-size and why i need max size also. and crop size why we do it becouse the ram or for the algorithem?
- my label now are in gray scale 0=bacground 100=myclass 255=ignorlabel there is another format that i need to do for my label and what id_to_label and label_to_id i need to create ?
thank alot
To my knowledge, id_to_label & label_to_id is only applied to the CityScape dataset. Here is the definition https://github.com/mcordts/cityscapesScripts/blob/master/cityscapesscripts/helpers/labels.py. You need do nothing to you label.
The crop size is used because the origin images are randomly re scaled [0.7, 1.3]. This method is used usually to prevent over-fitting.
thank you.
there is difference between background to ignore label?
how i choose the right crop size to my data ?
If you are doing pixel-level semantic labeling task, there is no difference between background and ignore label.
The crop size depends on your data size and the capacity of your model.
The author of this project choose 500*500.
This paper choose 713*713.
This paper choose 544*544 cropped from 800*800.
All the models run on CityScape dataset
I think that you can choose your crop size depending on the receptive field of your model. You can also choose a crop size by virtue of experience.
thank alot , just for be sure
its mean that my label image should be size hxw with value {0(background) , 1(myclass)} or {255(background) , 1(myclass)}
It looks like that you are doing binary segmentation. {0, 1} is better using Binary Cross Entropy Loss function. Wish it to be useful.
to replace softmax in this - mx.sym.softmax_cross_entropy?
there is somthing else that i need to notice?
No, I mean that Binary Cross Entropy Loss is better if you create your own model. There is no problem using this model directly but changing the class number to 2.
ok thank you