Feat: add `transform_to_rotated_box` and `box_xyxy_to_xyxyr` by DerrickUnleashed · Pull Request #342 · mlverse/torchvision

DerrickUnleashed · 2026-07-10T16:06:22Z

Summary

item_transform_bbox_rotate() converts image_with_bounding_box items to image_with_rotated_box with $boxes in xyxyr format. draw_bounding_boxes.image_with_rotated_box() renders them as rotated polygons.

Changes

R/ops-box_convert.R — box_xyxy_to_xyxyr(boxes, angle) now rotates boxes around their center by the given angle (radians) and returns the enclosing axis-aligned box + angle column.

R/target_transform_detection.R — item_transform_bbox_rotate(x, angle) S3 generic with methods for image_with_bounding_box and dataset.

R/vision_utils.R — new draw_bounding_boxes.image_with_rotated_box S3 method that draws rotated rectangles as polygons via graphics::polygon().

tests/testthat/test-ops-box_convert.R — 5 new tests for rotation (pi/2, pi/4, per-box angles, edge cases).

tests/testthat/test-target_transform_detection.R — 6 new tests for the item transform (basic, field preservation, empty, multiple, non-mutation, non-zero angle).

Closes #338

Convert axis-aligned xyxy bounding boxes to xyxyr format (with 0 rotation) for rotated object detection pipelines. Includes an item-level transform that works on image_with_bounding_box objects and detection datasets.

…to_xyxyr - Fix broadcasting bug: cx/cy/hw/hh now reshaped to (n, 1) to match 2D angle - Fix dim arg: use dim=-1 (last axis) instead of dim=2 for min/max on corners - Fix dimension mismatch in torch_cat: reshape min/max results back to (n, 1) - Use torch_cat instead of torch_stack for corners to avoid 3D tensor - Handle empty boxes edge case (n=0) before any operations - Handle scalar angle broadcast to multiple boxes via expand - Accept per-box angle tensors

Draws rotated bounding boxes as polygons instead of axis-aligned rects. Computes the 4 corners of the rotated rectangle from xyxyr format and draws them with graphics::polygon.

DerrickUnleashed · 2026-07-10T18:09:39Z

url <- "https://upload.wikimedia.org/wikipedia/commons/9/9a/Aeroflot_Airbus_A330_Kustov.jpg"

img <- base_loader(url) |>
  transform_to_tensor() |>
  transform_resize(c(300, 500))

item <- list(
  x = img,
  y = list(
    boxes = torch_tensor(matrix(c(40, 95, 475, 180), ncol = 4), dtype = torch_float32()),
    labels = "airplane",
    image_height = 300L,
    image_width = 500L
  )
)
class(item) <- c("image_with_bounding_box", "list")

angle_deg <- 355  
angle_rad <- angle_deg * pi / 180

item_rot <- item_transform_bbox_rotate(item, angle = angle_rad)

before <- draw_bounding_boxes(item, colors = "blue", width = 4)
after <- draw_bounding_boxes(item_rot, colors = "red", width = 4)

grid <- vision_make_grid(torch_stack(list(before, after), dim = 1)$to(dtype = torch_float32()))
tensor_image_browse(grid)

img <- torch_ones(3, 400, 400)  # white canvas

angles_deg <- seq(0, 340, by = 20)
angles_rad <- angles_deg * pi / 180
colors <- grDevices::hcl.colors(length(angles_rad), rev = TRUE)

base_box <- torch_tensor(matrix(c(180, 50, 220, 350), ncol = 4), dtype = torch_float32())

all_boxes <- torch_cat(lapply(angles_rad, function(a) {
  box_xyxy_to_xyxyr(base_box, angle = a)
}), dim = 1L)

all_labels <- paste0(angles_deg)

item <- list(
  x = img,
  y = list(
    boxes = all_boxes,
    labels = all_labels,
    image_height = 400L,
    image_width = 400L
  )
)
class(item) <- c("image_with_rotated_box", "list")

result <- draw_bounding_boxes(item, colors = colors, width = 2)
tensor_image_browse(result)

DerrickUnleashed · 2026-07-10T18:20:14Z

@cregouby , I just wanted to know if I'm going in the right direction.

cregouby · 2026-07-11T10:45:16Z

Hello @DerrickUnleashed,

This is the good direction, indeed, as a step toward implementing pytorch v2 transform.
Few remarks:
todo the spirit of item_transform_* is to have an item as input but also an item as output, because it will modify both $x and $y. So item_transform_bbox_rotate() should be named item_transform_rotate() and shall rotate both image and bboxes (like in the pytoch/vision vignette).
todo I made a mistake in the angle parameter : pytorch vision expect the angle in degrees (which is more accessible to human understanding) so could we switch back to angle in degrees, please ?
suggestion code organization We shouldn't mix target_transform and item_transform, so I would move all item_transform_bbox_* into a dedicated item_transforms_detection.R file
thought we may also using file organization that mimic the pytorch/vision code organization for v2 transforms and thus name it item_transform_geometry.R. Up to you to decide.

…egrees Align with pytorch/vision convention where angles are expressed in degrees for better human readability. Internally converts to radians for trigonometric computations.

…rendering The 5th column of xyxyr boxes is now in degrees after the preceding refactor. Convert to radians internally for trigonometric rendering.

…mage and boxes Rename item_transform_bbox_rotate() to item_transform_rotate() and move to a dedicated item_transform_geometry.R file (aligning with pytorch/vision file organization for geometric transforms). The function now rotates both the image (via transform_rotate) and the bounding boxes (via box_xyxy_to_xyr) together, matching the spirit of item_transform_* which should modify both and . Supports canvas expansion via the expand parameter. Angle parameter is in degrees (aligning with pytorch/vision convention).

…n rotation tests

The function converts bounding boxes to xyxyr format without modifying the image tensor, matching the original item_transform_bbox_rotate behavior.

DerrickUnleashed · 2026-07-11T16:18:06Z

url <- "https://upload.wikimedia.org/wikipedia/commons/9/9a/Aeroflot_Airbus_A330_Kustov.jpg"

img <- base_loader(url) |>
  transform_to_tensor() |>
  transform_resize(c(300, 500))

item <- list(
  x = img,
  y = list(
    boxes = torch_tensor(matrix(c(40, 95, 475, 180), ncol = 4), dtype = torch_float32()),
    labels = "airplane",
    image_height = 300L,
    image_width = 500L
  )
)
class(item) <- c("image_with_bounding_box", "list")

item_rot <- item_transform_rotate(item, angle = 355)

before <- draw_bounding_boxes(item, colors = "blue", width = 4)
after <- draw_bounding_boxes(item_rot, colors = "red", width = 4)

grid <- vision_make_grid(torch_stack(list(before, after), dim = 1)$to(dtype = torch_float32()))
tensor_image_browse(grid)

img <- torch_ones(3, 400, 400)  # white canvas

angles_deg <- seq(0, 340, by = 5)
colors <- grDevices::hcl.colors(length(angles_deg), rev = TRUE)

base_box <- torch_tensor(matrix(c(180, 50, 220, 350), ncol = 4), dtype = torch_float32())

all_boxes <- torch_cat(lapply(angles_deg, function(a) {
  box_xyxy_to_xyxyr(base_box, angle = a)
}), dim = 1L)

all_labels <- paste0(angles_deg)

item <- list(
  x = img,
  y = list(
    boxes = all_boxes,
    labels = all_labels,
    image_height = 400L,
    image_width = 400L
  )
)
class(item) <- c("image_with_rotated_box", "list")

result <- draw_bounding_boxes(item, colors = colors, width = 2)
tensor_image_browse(result)

cregouby

todo some small chrore

cregouby

todo Sorry I missed that : transform is intended for dataset augmentation, and shall rotate the image and the bbox, together, without bbox expansion.
This mean we have to use an additionnal box_rotate function and not the (to-be-kept) box_xyxy_to_xyxyr() function to rotate the box.

cregouby · 2026-07-12T07:46:55Z

+  expect_equal_to_r(result$y$boxes[1, 5], 0)
+})
+
+test_that("item_transform_rotate does not modify the image", {


todo we DO expect the image to be modified !

cregouby · 2026-07-12T07:47:41Z

+  expect_equal(torch::as_array(result$y$boxes[, 5]), c(0, 0, 0))
+})
+
+test_that("item_transform_rotate does not mutate input", {


todo we DO expect the image to be modified !

cregouby · 2026-07-12T08:06:16Z

+#' \eqn{r} is the rotation angle in degrees (counter-clockwise). The image
+#' is left unchanged. For axis-aligned boxes, \eqn{r = 0}.


todo the image shall be rotated as well (see https://docs.pytorch.org/vision/main/auto_examples/transforms/plot_rotated_box_transforms.html)

cregouby · 2026-07-12T08:07:39Z

+
+#' @export
+item_transform_rotate.image_with_bounding_box <- function(x, angle = 0) {
+  x$y$boxes <- box_xyxy_to_xyxyr(x$y$boxes, angle = angle)


question I guess the required rotation of the box is much simpler here, as there is no need to take the axis-aligned bounding box of the rotated corners..

DerrickUnleashed · 2026-07-12T11:57:36Z

Apologies for the oversight I seem to have missed that the image is also modified that is rotated will implment this and update tests accordingly

Hey, @cregouby , I have some questions to make sure the implementation is alright.

https://developer.nvidia.com/blog/detecting-rotated-objects-using-the-odtk/ -> this just rotates the bbox

The current task is just to rotate the image ? Do we need to rotate the bbox relative to the object.

If we are gonna go by using PyTorch docs then do we utilise the cxcywhr format instead ?

cregouby · 2026-07-13T19:39:09Z

I guess both features are valid: The one you've made could be renamed `target_transform_rotate_box()` (as it does not require the image) and the one with image and box rotation `item_transform_rotate()`. I guess you should focus more on the pytorch/vision/v2/transform for inspiration than from the Nvidia one (my bad) Le 12 juillet 2026 17:03:26 GMT+02:00, Derrick Richard ***@***.***> a écrit :

…

DerrickUnleashed left a comment (mlverse/torchvision#342) Apologies for the oversight I seem to have missed that the image is also modified that is rotated will implment this and update tests accordingly Hey, @cregouby , I have some questions to make sure the implementation is alright. https://developer.nvidia.com/blog/detecting-rotated-objects-using-the-odtk/ -> this just rotates the bbox The current task is just to rotate the image ? Do we need to rotate the bbox relative to the object. -- Reply to this email directly or view it on GitHub: #342 (comment) You are receiving this because you were mentioned. Message ID: ***@***.***>

DerrickUnleashed added 5 commits July 10, 2026 21:34

feat: add transform_to_rotated_box and box_xyxy_to_xyxyr

f6254ee

Convert axis-aligned xyxy bounding boxes to xyxyr format (with 0 rotation) for rotated object detection pipelines. Includes an item-level transform that works on image_with_bounding_box objects and detection datasets.

rename transform_to_rotated_box to item_transform_bbox_rotate

6bca06c

add tests for box_xyxy_to_xyxyr and item_transform_bbox_rotate

faf3ad3

feat: add draw_bounding_boxes.image_with_rotated_box S3 method

162db5c

Draws rotated bounding boxes as polygons instead of axis-aligned rects. Computes the 4 corners of the rotated rectangle from xyxyr format and draws them with graphics::polygon.

DerrickUnleashed closed this Jul 10, 2026

DerrickUnleashed reopened this Jul 10, 2026

DerrickUnleashed marked this pull request as ready for review July 10, 2026 18:19

DerrickUnleashed added 2 commits July 11, 2026 01:34

Update NEWS.md

4499582

Trigger Workflow

5ec69dc

cregouby reviewed Jul 11, 2026

View reviewed changes

Comment thread tests/testthat/test-ops-box_convert.R Outdated

cregouby requested changes Jul 11, 2026

View reviewed changes

Comment thread tests/testthat/test-ops-box_convert.R Outdated

cregouby mentioned this pull request Jul 11, 2026

Improve item_transform_ coverage against pytorch / vision /v2 / transforms #343

Open

cregouby and others added 9 commits July 11, 2026 13:55

separate item transforms in documentation

dc53c31

refactor(box_xyxy_to_xyxyr): switch angle parameter from radians to d…

d119921

…egrees Align with pytorch/vision convention where angles are expressed in degrees for better human readability. Internally converts to radians for trigonometric computations.

fix(draw_bounding_boxes): convert degrees to radians for rotated box …

637b5be

…rendering The 5th column of xyxyr boxes is now in degrees after the preceding refactor. Convert to radians internally for trigonometric rendering.

docs: update NEWS.md for item_transform_rotate refactor

e2af4f3

test: use expect_equal_to_r() instead of expect_equal(as.numeric()) i…

a7294a2

…n rotation tests

fix: add item_transform_geometry.R to Collate field in DESCRIPTION

65f5e04

revert(item_transform_rotate): only rotate bboxes, not the image

fc3baa7

The function converts bounding boxes to xyxyr format without modifying the image tensor, matching the original item_transform_bbox_rotate behavior.

docs: use airplane image example in item_transform_rotate docs

28e15bb

DerrickUnleashed requested a review from cregouby July 11, 2026 17:54

cregouby requested changes Jul 11, 2026

View reviewed changes

cregouby reviewed Jul 12, 2026

View reviewed changes

utilise test functions

6d12659

DerrickUnleashed added 2 commits July 12, 2026 17:24

Moving item_transform_geometry tests into a new file

d3571da

Renaming file according to conventions

bf42f67

		#' \eqn{r} is the rotation angle in degrees (counter-clockwise). The image
		#' is left unchanged. For axis-aligned boxes, \eqn{r = 0}.

Uh oh!

Conversation

DerrickUnleashed commented Jul 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Changes

Uh oh!

DerrickUnleashed commented Jul 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

DerrickUnleashed commented Jul 10, 2026

Uh oh!

cregouby commented Jul 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

DerrickUnleashed commented Jul 11, 2026

Uh oh!

cregouby left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

cregouby left a comment

Choose a reason for hiding this comment

Uh oh!

cregouby Jul 12, 2026

Choose a reason for hiding this comment

Uh oh!

cregouby Jul 12, 2026

Choose a reason for hiding this comment

Uh oh!

cregouby Jul 12, 2026

Choose a reason for hiding this comment

Uh oh!

cregouby Jul 12, 2026

Choose a reason for hiding this comment

Uh oh!

DerrickUnleashed commented Jul 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cregouby commented Jul 13, 2026 via email

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

DerrickUnleashed commented Jul 10, 2026 •

edited

Loading

DerrickUnleashed commented Jul 10, 2026 •

edited

Loading

cregouby commented Jul 11, 2026 •

edited

Loading

DerrickUnleashed commented Jul 12, 2026 •

edited

Loading