Mask_rcnn: Does data augmentation works for VGG Segmentations ? (model.train, augmentation=augmentation)

[nayzen]

Hi everyone!
Firstly, thank you for this repo, it is amazing.
I have 134 classes (for about 2800 photos --- 21 photos / classe). I used VGG image annotator tool to do my annotations (2800 segmentations...)

I have quite good results bu I would like to augment my database using the augmentation parameter below (in model.train) if the segmentations will be augmented too ?

My segmentations are coordinates in a .JSON file and I would like to do like the Imgaug example below:

Except if the augmentation parameter already does what I want..
Waiting for your help.. Thanks !

[bilgan]

Yes, it does what you want.

[eyildiz-ugoe]

I try to do the same, but I get:

Traceback (most recent call last):
  File "script.py", line 539, in <module>
    train(model)
  File "script.py", line 248, in train
    layers='heads', augmentation=augmentation)
NameError: name 'augmentation' is not defined

Does anyone know why it doesn't work?

[nayzen]

@eyildiz-ugoe You need to define "augmentation" before. Like in this example :

Regards,
Antoine

[nikhil-occipitaltech]

@nayzen Hey man, your solution works. But how exactly, is something I am struggling to grasp. Augmentating the images is not that big of a deal but how excatly the related json with the the contour points is getting modified?

[nayzen]

Hey @nikhil-occipitaltech , I don't really know how the contour points is getting modified and I'm interested in finding out as well. But it seems that the imgaug library does automatically this task.

[nikhil-occipitaltech]

@nayzen I am trusting it to be correct.