Any plan about data augmentation

[zhiqwang]

❓ Is there any plan about data augmentation

Hi, DETR teams,

According to the implementation in ultralytics/yolov3#310 (comment), and a similar are discussed in AlexeyAB/darknet#3114 (comment). It seems that such augmentation such as the mosaic techniques is helpful to detect smaller size object. I quote Jocher's conclusions below.

The smaller cars are detected earlier with less blinking and cars of all sizes show better behaved bounding boxes.

I check make_coco_transforms of this repo, and visualized the augmented images and labels in
VOC dataset (use the same config of make_coco_transforms here). Since the utilization of RandomSizeCrop, all the labels associated to an image may be cropped. (So this repo supports training with no targets in an image? 🤔️)

I want to know whether is there some plan about data augmentation.

Thank you!

[fmassa]

Hi,

The current implementation supports a number of data augmentations, including random resizing and random crops.

If I understand it correctly, the mozaic data augmentation that you pointed out is very similar (if not equivalent) to doing random resize on each image. Indeed, in the same mini-batch, we can have images of different resolutions, and it can be customized, so that a very large image could be in the same minibatch as a very small image.

Since the utilization of RandomSizeCrop, all the labels associated to an image may be cropped. (So this repo supports training with no targets in an image? 🤔️)

Yes, that's correct, we support feeding training images without any annotations (which generally happens with random crops).

I believe I've answered your questions, and as such I'm closing the issue, but let us know if you have further questions.