search

david8862 / keras-YOLOv3-model-set

end-to-end YOLOv4/v3/v2 object detection pipeline, implemented on tf.keras with different technologies
MIT License
640 stars 222 forks source link

Data Augmentation #158

Open enliden1 opened 3 years ago

enliden1 commented 3 years ago

I’m training a yolov3 model with a small dataset (2000 images). Do you have any recommandation how to add augmentation to “increase” the dataset? I’m already implementing Mosaic data augmentation, with little improvement. I guess i could increase the probability for the random_vertical_flip, random_rotate etc. functions in common/data_utils.py, do you have any other recommendation?

Thank you for sharing this great project!

david8862 commented 3 years ago

@enliden1 many thanks. Choose of augmentation may depends on your object type or user scenario, but generally the color related (chroma/contrast/sharpness, etc.) should be safe. And gridmask may also worth to have a try.

enliden1 commented 3 years ago

It looks like gridmask is already implemented in the code (yolo3/data.py line 79):

image, boxes = random_gridmask(image, boxes)