Google Brain just published a new paper that shows a new ‘best policy’ for leveraging data augmentation specific to object detection is a key aspect for training and show new state of the art results with it.
Specifically they find that simply doing the same augmentations as those for image classification is sub-optimal. They setup a RL platform and tested out a whole slew of augmentations and produced a new augmentation policy for training.
The top 3 augmentations were: Rotation / Equalize / Shift bounding box Y axis
These are the most frequently used in their learned policy for training and they show state of the art results using it with RetinaNet and COCO dataset.
I wrote an article summarizing their findings (including their findings that data augmentation enforces it’s own regularization):
and highly recommend reading the paper in it’s entirety if you have time:
They have the source code for their optimal policy and augmentations, but it’s in TensorFlow.
Anyway, really nice info to know if you are working on object detection and want to optimize your training results!