facebookresearch / detectron2

Detectron2 is a platform for object detection, segmentation and other visual recognition tasks.
https://detectron2.readthedocs.io/en/latest/
Apache License 2.0
30.46k stars 7.48k forks source link

Image sharpness data augmentation #2394

Open gauravsinghal-oviva opened 3 years ago

gauravsinghal-oviva commented 3 years ago

🚀 Feature

I have been using Detectron2 for 3 months. I have tried nearly all the data augmentation techniques which are provided out-of-the-box by Detectron2. I would like to propose a new data augmentation technique that will enhance the input image quality, "Sharpness".

Motivation & Examples

Image sharpness enhances the edges of objects in the image which improves the masking and overall loss. A lot of people use sharped images as a separate dataset and put them in the training, which increases the training images. With this new augmentation, selective images will get sharped and seen by the model.

bpfliegel commented 3 years ago

You can already statically preaugment your dataset like you described above. A better technique is however, to include dynamic augmentations in the training pipeline.

Please refer to the guide at: https://detectron2.readthedocs.io/tutorials/data_loading.html

You can use e.g. the albumentations library for sharpness or other augmentations: https://github.com/albumentations-team/albumentations

Nwaf-sl commented 3 years ago

How to enhance and expand the data set and increase the number of training sets? Dynamic online enhancement can not increase the number