Compose multiple transforms on image and annotation.

Randomly choose a size from sizes to resize the image and boxes

Resize the image for the shortest edge to be a fixed size(size). If longest edge is longer than max_size, than resize the image for the lon

Color jitter, keep annotation

Gaussian blur augmentation in SimCLR https://arxiv.org/abs/2002.05709 Adapted from MoCo: https://github.com/facebookresearch/moco/blob

Normalize image with mean and std and convert box from [x, y, x, y] to [cx, cy, w, h]

Random grayscale, keep annotation

Random horizontal flip. When doing flip, flip the boxes at the same time.

Convert PIL image to Tensor and keep annotation.

This is a more standard version of the position embedding, very similar to the one used by the Attention is all you need paper, generalized

Randomly selects between transforms1 and transforms2, with probability p for transforms1 and (1 - p) for transforms2

Absolute pos embedding, learned.

Random erasing, keep annotation

Convert Tensor to PIL image and keep annotation.