search

facebookresearch / Detectron

FAIR's research platform for object detection research, implementing popular algorithms like Mask R-CNN and RetinaNet.
Apache License 2.0
26.21k stars 5.45k forks source link

Detectron2 In-Place One-to-Many Augmentations #1023

Open kmeng01 opened 3 years ago

kmeng01 commented 3 years ago

Hi there! This is a simple question: I haven't been able to find documentation for applying in-place augmentations that can produce multiple transformed "images" from a single image. It appears that the current system simply chooses some augmentation, applies it to the image, and moves onto the next one.

Am I missing something, and if not, is there any way this feature can be implemented with Detectron? Thank you!

joshuastobaugh57 commented 3 years ago

Great qestion

On Friday, May 14, 2021, Kevin Meng @.***> wrote:

Hi there! This is a simple question: I haven't been able to find documentation for applying in-place augmentations that can produce multiple transformed "images" from a single image. It appears that the current system simply chooses some augmentation, applies it to the image, and moves onto the next one.

Am I missing something, and if not, is there any way can be implemented with Detectron? Thank you!

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/facebookresearch/Detectron/issues/1023, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADMVZR54HT72XNVRIQ545ELTNXPCVANCNFSM445NJEAQ .

-- Sent from Gmail Mobile

SamuelllG commented 1 year ago

You can implement your own dataset mapper and overwrite a few classes, e.g. sampler or dataset.