[FEATURE] Is there any way to apply "the same" transformation on two images?

[sangrockEG]

Issue Description:

Sometimes, we need to apply consistent transformations to multiple images, particularly in tasks like semantic segmentation. In such cases, both the input image and its corresponding GT map (which possibly has more or less than three channels) require identical augmentation. However, existing transformations are tailored for a single three-channel RGB PIL image, leading to the need for complex random number generator (RNG) control and manually adjusting the number of channels, to ensure consistent augmentation across the image and GT map.

To simplify this process, it would be advantageous to extend geometrical augmentations—such as random cropping or shearing—to accommodate multiple maps as supplementary inputs. Of course, these maps should share the same dimensions (height and width) as the input image but can have arbitrary numbers of channel. Consequently, all maps can undergo augmentation in unison with the input image.

Furthermore, addressing scenarios involving combined photometric and geometrical augmentations, such as auto_augmentation, is essential. When inputting an image and corresponding maps, geometrical augmentations should apply to both images and maps, while photometric augmentations should affect only the image.

While I'm not certain about the technical difficulty of adding this feature, it has the potential to greatly increase the convenience of the augmentation process.

Proposed Changes:

Extended Geometrical Augmentations: Modify the augmentation process to support multiple maps in addition to the input image. Maps should share the same dimensions as the input image, while the number of channels can vary.

Mixed Augmentations Handling: Implement support for scenarios involving both photometric and geometrical augmentations, such as auto_augmentation. Apply geometrical augmentations to both images and maps, while photometric augmentations should be limited to the input image.

Benefits:

Simplified Workflow: Eliminate the need for intricate RNG control when applying consistent transformations to multiple images and maps.

Enhanced Data Consistency: Geometrical augmentations will be synchronized across images and maps, maintaining data integrity for tasks like semantic segmentation.

Flexibility: Support various map channel configurations, accommodating diverse applications beyond RGB images.