Detectron2: how to make full use of data augmentation

[AaronXinWu]

❓ How to use Detectron2

hello team, Can you answer one queation for a novice:

1. what's the parameter format for data augmentation function in 'config.yaml '. I can only find :
   '''
   INPUT:
   CROP:
   ENABLED: false
   SIZE:
   
   
   
   • 0.9
   
   
   • 0.9
     
     TYPE: relative_range
     
     FORMAT: BGR
     
     MASK_FORMAT: polygon
     
     MAX_SIZE_TEST: 1333
     
     MAX_SIZE_TRAIN: 1333
     
     MIN_SIZE_TEST: 800
     
     MIN_SIZE_TRAIN:
   
   
   • 640
   
   
   • 672
   
   
   • 704
   
   
   • 736
   
   
   • 768
   
   
   • 800
     
     MIN_SIZE_TRAIN_SAMPLING: choice
     
     ...
     
     '''
     
     but, I find some other function in "transform_gen.py"
     
     '''
   
   

"RandomBrightness",

24 | "RandomContrast",
25 | "RandomCrop",
26 | "RandomExtent",
27 | "RandomFlip",
28 | "RandomSaturation"
'''

[ppwwyyxx]

> "RandomBrightness",

24 | "RandomContrast",
25 | "RandomCrop",
26 | "RandomExtent",
27 | "RandomFlip",
28 | "RandomSaturation"

These are augmentation utilities you can call if you use your custom data loader.

[AaronXinWu]

"RandomBrightness",

24 | "RandomContrast",
25 | "RandomCrop",
26 | "RandomExtent",
27 | "RandomFlip",
28 | "RandomSaturation"

These are augmentation utilities you can call if you use your custom data loader.

Thank you for your kind reply.
Do you mean I need to comprehend the whole project and modify the code to build my own data loader.
Is there any interfaces or guidance.

[xmyqsh]

In some cases, data augmentation also not works.
For example, the training and validation set from the new style map based dataset have some don't care area, but you have not masked this area manually.
It is the duty of the dataset user(me) to mask this don't care area which may
have lots of unannotated objects before training.

AaronXinWu [e-mail] 于2019年12月14日周六 下午4:59写道：

"RandomBrightness",

24 | "RandomContrast",
25 | "RandomCrop",
26 | "RandomExtent",
27 | "RandomFlip",
28 | "RandomSaturation"

These are augmentation utilities you can call if you use your custom data
loader.

Thank you for your kind reply.
Do you mean I need to comprehend the whole project and modify the code to
build my own data loader.
Is there any interfaces or guidance.

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
https://github.com/facebookresearch/detectron2/issues/527?email_source=notifications&email_token=AD2E4TF3O4EPNISRQNQ6I5DQYSN53A5CNFSM4J2Y5FS2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEG357SA#issuecomment-565698504,
or unsubscribe
https://github.com/notifications/unsubscribe-auth/AD2E4TE3W4S6A7GM36BAZCDQYSN53ANCNFSM4J2Y5FSQ
.

[ppwwyyxx]

https://detectron2.readthedocs.io/tutorials/data_loading.html explains how to use your own data loader. You need a custom data mapping function to add new augmentations.
One example of such customization is in https://github.com/facebookresearch/detectron2/blob/22b70a8078eb09da38d0fefa130d0f537562bebc/projects/DensePose/train_net.py#L37-L38

[AaronXinWu]

thank you for your support!!