Feat: add item_transform_random_erasing for detection and segmentation items

[DerrickUnleashed]

Adds item_transform_random_erasing() for randomly erasing a rectangular region of a dataset item, covering object detection items/datasets, semantic segmentation items/datasets, and rotated-box items.

Currently only the image pixels are erased; boxes and masks are left unchanged, matching PyTorch's documented RandomErasing behaviour.

Question: Should we switch to dropping bounding boxes whose centre falls inside the erased region?

Closes #358

url <- "https://upload.wikimedia.org/wikipedia/commons/b/b6/Felis_catus-cat_on_snow.jpg"

# ========== 1. DETECTION ITEM ==========
img <- base_loader(url) |> transform_to_tensor()
boxes <- torch_tensor(matrix(c(600, 200, 2880, 1860), ncol = 4), dtype = torch_float32())
det_item <- list(x = img, y = list(boxes = boxes, labels = "CAT"))
class(det_item) <- c("image_with_bounding_box", "list")

det_erased <- item_transform_random_erasing(det_item)

p1 <- draw_bounding_boxes(det_item, colors = "blue", width = 10)$to(torch_float())$div(255)
p2 <- draw_bounding_boxes(det_erased, colors = "red", width = 10)$to(torch_float())$div(255)
grid1 <- vision_make_grid(torch_stack(list(p1, p2)), scale = TRUE)
tensor_image_browse(grid1)

# ========== 2. DETECTION DATASET ==========
ds_det <- pascal_detection_dataset(year = "2007", split = "trainval",
                                   transform = transform_to_tensor, download = TRUE)
orig_det <- ds_det[1]
ds_det_erased <- item_transform_random_erasing(ds_det)
erase_det <- ds_det_erased[1]

p3 <- draw_bounding_boxes(orig_det, colors = "blue", width = 5)$to(torch_float())$div(255)
p4 <- draw_bounding_boxes(erase_det, colors = "red", width = 5)$to(torch_float())$div(255)
grid2 <- vision_make_grid(torch_stack(list(p3, p4)), scale = TRUE)
tensor_image_browse(grid2)

# ========== 3. SEGMENTATION ITEM ==========
img2 <- base_loader(url) |> transform_to_tensor()
h <- img2$shape[2]; w <- img2$shape[3]

mask1 <- torch_zeros(h, w, dtype = torch_bool()); mask1[, 1:150] <- TRUE
mask2 <- torch_zeros(h, w, dtype = torch_bool()); mask2[, 350:500] <- TRUE
masks <- torch_stack(list(mask1, mask2))
seg_item <- list(x = img2, y = list(masks = masks, labels = torch_tensor(c(1L, 2L)),
                                    image_height = h, image_width = w))
class(seg_item) <- c("image_with_segmentation_mask", "list")

seg_erased <- item_transform_random_erasing(seg_item)

p5 <- draw_segmentation_masks(seg_item, alpha = 0.5, colors = c("red", "blue"))$to(torch_float())$div(255)
p6 <- draw_segmentation_masks(seg_erased, alpha = 0.5, colors = c("red", "blue"))$to(torch_float())$div(255)
grid3 <- vision_make_grid(torch_stack(list(p5, p6)), scale = TRUE)
tensor_image_browse(grid3)

# ========== 4. SEGMENTATION DATASET ==========
ds_seg <- pascal_segmentation_dataset(year = "2007", split = "trainval",
                                      transform = transform_to_tensor, download = TRUE)
orig_seg <- ds_seg[1]
ds_seg_erased <- item_transform_random_erasing(ds_seg)
erase_seg <- ds_seg_erased[1]

p7 <- draw_segmentation_masks(orig_seg, alpha = 0.5)$to(torch_float())$div(255)
p8 <- draw_segmentation_masks(erase_seg, alpha = 0.5)$to(torch_float())$div(255)
grid4 <- vision_make_grid(torch_stack(list(p7, p8)), scale = TRUE)
tensor_image_browse(grid4)

DETECTION ITEM

DETECTION DATASET

SEGMENTATION ITEM

SEGMENTATION DATASET