Implement SAHI: (Slicing Aided Hyper Inference) as both `transform_` and `target_transform_`

[cregouby]

Current Situation

Large-scale and high-resolution images (2000 pts +) get poor result in object detection model as image resize to model input size remove a lot of details.

Expected situation

SAHI: (Slicing Aided Hyper Inference) shall be available as transform to turn a single large-scale image into a batch of sliced images.

API suggestion
transform_sahi_crop() and target_transform_sahi_crop_coco() for coco format annotations

[DerrickUnleashed]

I had a suggestion for this can we have the api as

target_transform_sahi_crop(
  y,
  crop_height,   #The height of each small box
  crop_width,    #The width of each small box
  stride_height = crop_height,    #The amount of distance moved down
  stride_width = crop_width     #The amount of distance moved right
)

I wanted to know your comments before proceeding @cregouby

[cregouby]

thought We should maintain consistency within torchvision::transfrom parameters, and at the same time allow end-user to switch back-and-forth to the reference documentation of sahi in https://obss.github.io/sahi/slicing/#sahi.slicing.slice_coco(min_area_ratio).

suggestion You can maintain / inherit parameter size from the rest of the transform_* function for the output size
suggestion I like the "overlap_ratio" term as it is transparent. why not having an overlap_size_ratio = c(0.2, 0.1) for this ?

[DerrickUnleashed]

thought We should maintain consistency within torchvision::transfrom parameters, and at the same time allow end-user to switch back-and-forth to the reference documentation of sahi in https://obss.github.io/sahi/slicing/#sahi.slicing.slice_coco(min_area_ratio).

suggestion You can maintain / inherit parameter size from the rest of the transform_* function for the output size suggestion I like the "overlap_ratio" term as it is transparent. why not having an overlap_size_ratio = c(0.2, 0.1) for this ?

This Sounds like a good idea

transform_sahi_crop(
  size = c(512, 512),
  overlap_size_ratio = c(0.2, 0.2)
)

target_transform_sahi_crop(
  y,
  crop_windows,
  min_area_ratio = 0.1
)