AlbumentationsX: One Augmentation Pipeline for Images and Related Annotations
The paper’s core point is that augmentation should sample randomness once, then apply it consistently across every image-linked annotation.
AlbumentationsX puts transforms, probabilities, annotation settings, and the random seed inside one `Compose` object.
That keeps crops, flips, masks, boxes, keypoints, stereo views, video frames, and volumes aligned instead of letting separate code paths drift.
It can save a pipeline definition, show a single augmentation call, and replay it.
The author notes that users still have to decide whether a transform preserves the correct label for their task.
ArXiv · AI/CL/LG's note
AlbumentationsX puts transforms, probabilities, annotation settings, and the random seed inside one `Compose` object.
That keeps crops, flips, masks, boxes, keypoints, stereo views, video frames, and volumes aligned instead of letting separate code paths drift.
It can save a pipeline definition, show a single augmentation call, and replay it.
The author notes that users still have to decide whether a transform preserves the correct label for their task.
ArXiv · AI/CL/LG's note
score 5