How to get the picture class_name? about object-localization HOT 3 CLOSED

Hi,

when you want to localize the dog/cat and also want to distinguish between the two classes, you have a mix of regression and classification.

This means:

1. Add to the csv file a column "class" and load it in train_model.py
2. the CNN has to output now 5 values (4 coords + 0 or 1 for class)
3. You might have to use a merge layer https://keras.io/layers/merge/#concatenate_1 (one branch gets softmax, one branch gets identity activation i.e. "linear")
4. Write a custom loss function that calculates MSE and cross entropy. The coord loss will have a stronger influence, maybe MSE will work for the class too. You could also weight the losses differently.

from object-localization.

Thank you for your advice, it is good for me.Good day!

from object-localization.

I just pushed some code that implements this (example_2).

from object-localization.