Add target_transform_resize() as object detection transform

DESCRIPTION

@@ -104,6 +104,7 @@ Collate:
     'models-vit.R'
     'ops-box_convert.R'
     'ops-boxes.R'
+    'target_transform_detection.R'
     'transforms-array.R'
     'transforms-defaults.R'
     'transforms-generics.R'

NAMESPACE

@@ -7,6 +7,7 @@ S3method(draw_segmentation_masks,default)
 S3method(draw_segmentation_masks,image_with_segmentation_mask)
 S3method(draw_segmentation_masks,torch_tensor)
 S3method(get_image_size,"magick-image")
+S3method(get_image_size,torch_tensor)
 S3method(transform_adjust_brightness,default)
 S3method(transform_adjust_brightness,torch_tensor)
 S3method(transform_adjust_contrast,default)
@@ -208,6 +209,7 @@ export(rf100_peixos_segmentation_dataset)
 export(rf100_underwater_collection)
 export(search_collection)
 export(target_transform_coco_masks)
+export(target_transform_resize)
 export(target_transform_trimap_masks)
 export(tensor_image_browse)
 export(tensor_image_display)

NEWS.md

@@ -4,6 +4,10 @@
 
 * Added `model_lw_detr_tiny()`, `model_lw_detr_small()`, `model_lw_detr_medium()` and `model_lw_detr_large()` for real-time object detection (@srishtiii28, #328).
 
+## New features
+
+* Added `target_transform_resize` to manage bounding_box resizing in parallel to `transform_resize` for images (#337).
+
 ## Bug fixes and improvements
 
 * `nms()` now uses `torchvisionlib::ops_nms()` when torchvisionlib is installed, speeding up inference for `model_fasterrcnn_*()` and `model_maskrcnn_*()` (#321, #322).

R/dataset-coco.R

@@ -147,16 +147,16 @@ coco_detection_dataset <- torch::dataset(
       boxes = boxes,
       labels = labels,
       area = area,
-      iscrowd = iscrowd
+      iscrowd = iscrowd,
+      image_height = height,
+      image_width = width
     )
 
     if (!is.null(self$transform)) {
       x <- self$transform(x)
     }
 
     if (!is.null(self$target_transform)) {
-      y$image_height <- height
-      y$image_width <- width
       y <- self$target_transform(y)
     }
 
@@ -298,16 +298,16 @@ coco_segmentation_dataset <- torch::dataset(
     y <- list(
       labels = labels,
       iscrowd = iscrowd,
-      segmentation = anns$segmentation
+      segmentation = anns$segmentation,
+      image_height = height,
+      image_width = width
     )
 
     if (!is.null(self$transform)) {
       x <- self$transform(x)
     }
 
     if (!is.null(self$target_transform)) {
-      y$image_height <- height
-      y$image_width <- width
       y <- self$target_transform(y)
     }
 

R/target_transform_detection.R

@@ -0,0 +1,89 @@
+#' Resize the bounding boxes of a detection target
+#'
+#' Adjusts the bounding box coordinates in a detection target to match a
+#'   resized image. The target is expected to contain a `boxes` field with
+#'   shape `(N, 4)` in xyxy format and an `orig_size` field with the original
+#'   image dimensions `(H, W)`.
+#'
+#' @param target A list containing at least:
+#'   \itemize{
+#'     \item `boxes` — tensor of shape `(N, 4)` with bounding boxes in xyxy format
+#'     \item `orig_size` — integer vector or tensor of length 2 with `(height, width)`
+#'     \item Other fields (labels, image_id, etc.) are preserved unchanged
+#'   }
+#' @param size Desired output size. If `size` is a integer vector of length 2
+#'   like `c(h, w)`, output size will be matched to this. If `size` is a bare integer,
+#'   smaller edge of the image will be matched to this number.
+#'   i.e, if height > width, then image will be rescaled to
+#'   `(size * height / width, size)`.
+#'
+#' @return A list with the same structure as the input target, where:
+#'   \itemize{
+#'     \item `boxes` are rescaled according to the resize factors
+#'     \item `orig_size` is updated to the new dimensions
+#'   }
+#'
+#' @details
+#' This function should be composed with a [transform_resize()] in the
+#'   `transform` pipeline having the same `size` to ensure resized bounding boxes
+#'  remain aligned with the resized image.
+#'
+#' @examples
+#' \dontrun{
+#' target <- list(
+#'   boxes = torch_tensor(matrix(c(10, 20, 50, 60), ncol = 4)),
+#'   labels = torch_tensor(1L, dtype = torch_long()),
+#'   image_height  = 100L,
+#'   image_width = 100L)
+#' )
+#'
+#' # Resize to fixed size
+#' transform_fn <- target_transform_resize(c(200L, 200L))
+#' new_target <- transform_fn(target)
+#'
+#' # Resize with proportional scaling
+#' transform_fn <- target_transform_resize(50L)
+#' new_target <- transform_fn(target)
+#' }
+#'
+#' @family target_transforms_detection
+#'
+#' @export
+target_transform_resize <- function(target, size) {
+  if (!"image_height" %in% names(target) || !"image_width" %in% names(target)) {
+    cli::cli_abort("Target must contain both 'image_height' and 'image_width' field")
+  }
+  # Extract original dimensions
+  orig_h <- target$image_height
+  orig_w <- target$image_width
+
+  # Compute new dimensions
+  if (length(size) == 1L) {
+    # Proportional resize: match smaller edge to size
+    scale <- size / max(orig_h, orig_w)
+    new_h <- round(orig_h * scale)
+    new_w <- round(orig_w * scale)
+  } else {
+    # Fixed size resize
+    c(new_h, new_w) %<-% size
+  }
+
+  # Compute scale factors (reuse orig_h / orig_w)
+  scale_h <- new_h / orig_h
+  scale_w <- new_w / orig_w
+
+  # Resize bounding boxes (xyxy format)
+  boxes <- target$boxes$clone()
+  boxes_resized <- boxes
+  boxes_resized[, 1L] <- boxes[, 1L] * scale_w # xmin
+  boxes_resized[, 2L] <- boxes[, 2L] * scale_h # ymin
+  boxes_resized[, 3L] <- boxes[, 3L] * scale_w # xmax
+  boxes_resized[, 4L] <- boxes[, 4L] * scale_h # ymax
+
+  # Update target
+  target$boxes <- boxes_resized
+  target$image_height <- new_h
+  target$image_width <- new_w
+
+  target
+}