diff --git a/DESCRIPTION b/DESCRIPTION
index ead8fbd6..16041ee4 100644
--- a/DESCRIPTION
+++ b/DESCRIPTION
@@ -118,7 +118,7 @@ Collate:
     'transforms-magick.R'
     'transforms-tensor.R'
     'utils.R'
-    'vision_utils.R'
+    'vision-utils.R'
 Depends: 
     R (>= 3.5)
 LazyData: true
diff --git a/NEWS.md b/NEWS.md
index 2a1f7c25..21e8c408 100644
--- a/NEWS.md
+++ b/NEWS.md
@@ -36,9 +36,11 @@
 * Detection datasets (`coco_detection_dataset()`, `pascal_detection_dataset()` and the `rf100_*_collection()`s) now inherit the `object_detection_dataset` class and their item target `y` the `object_detection_target` class.
   Segmentation datasets (`coco_segmentation_dataset()`, `pascal_segmentation_dataset()`, `cityscapes_dataset()`, `oxfordiiitpet_segmentation_dataset()` and `rf100_peixos_segmentation_dataset()`) now inherit the `segmentation_dataset` and `segmentation_target` classes. 
   Target transforms now dispatch on those classes instead of inspecting the target fields: `target_transform_resize()`, `target_transform_rotate()`, `target_transform_affine()` and `target_transform_sahi_crop()` take an `object_detection_target`, and `target_transform_coco_masks()` and `target_transform_trimap_masks()` a `segmentation_target`. A bare list is no longer accepted as a target, so a hand-built one needs its class set (@srishtiii28, #391).
+* Added a "Visualization utilities" article covering `vision_make_grid()`, `draw_bounding_boxes()`, `draw_segmentation_masks()` and `draw_keypoints()` on the output of `model_rfdetr_base()` and `model_fcn_resnet50()` (@srishtiii28, #400).
 
 ## Bug fixes and improvements
 
+* `draw_bounding_boxes()` now accepts degenerated bounding-box by default with the `lazy = TRUE` parameter (#400).
 * `vision_make_grid()` now accepts multiple 3D tensors with mixed uint8 and float dtype (#398).
 * `transform_random_affine()` now accepts a bare number for `shear`. It used to widen `degrees`
   instead of `shear`, which left the shear range incomplete and made the sampling fail (#390).
diff --git a/R/vision_utils.R b/R/vision-utils.R
similarity index 95%
rename from R/vision_utils.R
rename to R/vision-utils.R
index 0a926992..aa1be4d5 100644
--- a/R/vision_utils.R
+++ b/R/vision-utils.R
@@ -145,6 +145,28 @@ vision_make_grid.torch_tensor <- function(tensor, ..., scale = TRUE, per_row = 8
 }
 
 
+check_bbox_is_xyxy <- function(boxes, lazy = TRUE) {
+  valid <- (boxes[, 1] < boxes[, 3])$logical_and(boxes[, 2] < boxes[, 4])
+
+  if (lazy) {
+    if ((!valid)$any()$item()) {
+      boxes <- boxes[valid, ]
+    }
+  } else {
+    if ((!valid)$any()$item()) {
+      invalid_indices <- which(as.logical(!valid))
+      first_idx <- invalid_indices[1]
+      first_box <- as.numeric(boxes[first_idx, ])
+      cli_abort(c(
+        "Bounding box {.val {first_idx}} is not in valid xyxy format.",
+        "x" = "xmin ({.val {first_box[1]}}) must be < xmax ({.val {first_box[3]}}), and ymin ({.val {first_box[2]}}) must be < ymax ({.val {first_box[4]}})."
+      ))
+    }
+  }
+
+  boxes
+}
+
 #' Draws bounding boxes on image.
 #'
 #' Draws bounding boxes on top of one image tensor
@@ -166,6 +188,8 @@ vision_make_grid.torch_tensor <- function(tensor, ..., scale = TRUE, per_row = 8
 #' @param width  Width of text shift to the bounding box.
 #' @param font NULL for the current font family, or a character vector of length 2 for Hershey vector fonts.
 #' @param font_size The requested font size in points.
+#' @param lazy if `TRUE`, silently filter out degenerate bounding boxes (xmin >= xmax or ymin >= ymax).
+#'    If `FALSE`, error on the first non-conforming box.
 #' @param ... Additional arguments passed to methods.
 #'
 #' @return  torch_tensor of size (C, H, W) of dtype uint8: Image Tensor with bounding boxes plotted.
@@ -177,7 +201,7 @@ vision_make_grid.torch_tensor <- function(tensor, ..., scale = TRUE, per_row = 8
 #' x <- torch::torch_randint(low = 1, high = 160, size = c(12,1))
 #' y <- torch::torch_randint(low = 1, high = 260, size = c(12,1))
 #' boxes <- torch::torch_cat(c(x, y, x + 20, y +  10), dim = 2)
-#' bboxed <- draw_bounding_boxes(image_tensor, boxes, colors = "black", fill = TRUE)
+#' bboxed <- draw_bounding_boxes(image_tensor, boxes, colors = "black", label = "label", fill = TRUE)
 #' tensor_image_browse(bboxed)
 #' }
 #' }
@@ -202,7 +226,8 @@ draw_bounding_boxes.torch_tensor <- function(x,
                                              fill = FALSE,
                                              width = 1,
                                              font = c("serif", "plain"),
-                                             font_size = 10, ...) {
+                                             font_size = 10,
+                                             lazy = TRUE, ...) {
   rlang::check_installed("magick")
 
   if (x$ndim == 4 && x$size(1) == 1) x <- x$squeeze(1)
@@ -215,14 +240,13 @@ draw_bounding_boxes.torch_tensor <- function(x,
     x$permute(c(2, 3, 1))$to(device = "cpu") %>% as.array()
   } else type_error("`x` should be torch_uint8 or torch_float")
 
-  if (boxes$shape[2] == 4 &&
-      ((boxes[, 1] >= boxes[, 3])$any() %>% as.logical() ||
-       (boxes[, 2] >= boxes[, 4])$any() %>% as.logical())) {
-    value_error("Boxes must be in c(xmin, ymin, xmax, ymax) format")
-  }
+  boxes <- check_bbox_is_xyxy(boxes, lazy = lazy)
   num_boxes <- boxes$shape[1]
   if (num_boxes == 0) {
-    cli_warn("boxes doesn't contain any box. No box was drawn")
+    cli_warn(if (lazy)
+      "No valid bounding box to draw after filtering degenerate boxes."
+    else
+      "boxes doesn't contain any box. No box was drawn.")
     return(x)
   }
   if (!is.null(labels) && inherits(labels, "torch_tensor")) {
@@ -352,10 +376,12 @@ draw_bounding_boxes.image_with_rotated_box <- function(x,
                                                        fill = FALSE,
                                                        width = 1,
                                                        font = c("serif", "plain"),
-                                                       font_size = 10, ...) {
+                                                       font_size = 10,
+                                                       lazy = TRUE, ...) {
   rlang::check_installed("magick")
 
-  boxes <- x$y$boxes
+  boxes <- check_bbox_is_xyxy(x$y$boxes, lazy = lazy)
+
 
   img_to_draw <- if (x$x$dtype == torch_uint8()) {
     x$x$div(255)$permute(c(2, 3, 1))$to(device = "cpu") %>% as.array()
@@ -627,6 +653,7 @@ draw_segmentation_masks.torch_tensor <- function(x,
   if (num_masks %% length(colors) != 0) {
     cli_abort("colors vector of size {.value {length(colors)}} cannot be broadcasted on {.value {num_masks}} masks")
   }
+  colors <- rep_len(colors, num_masks)
 
   color_tt <- colors %>%
     grDevices::col2rgb() %>%
diff --git a/_pkgdown.yml b/_pkgdown.yml
index d359d280..fecd1650 100644
--- a/_pkgdown.yml
+++ b/_pkgdown.yml
@@ -31,6 +31,8 @@ navbar:
         href: articles/examples/fcnresnet.html
       - text: keypoints
         href: articles/examples/keypoints.html
+      - text: visualization-utilities
+        href: articles/examples/visualization-utilities.html
 
 reference:
 - title: Image Transforms
diff --git a/man/coco_polygon_to_mask.Rd b/man/coco_polygon_to_mask.Rd
index 6d81b11e..62e77f08 100644
--- a/man/coco_polygon_to_mask.Rd
+++ b/man/coco_polygon_to_mask.Rd
@@ -1,5 +1,5 @@
 % Generated by roxygen2: do not edit by hand
-% Please edit documentation in R/vision_utils.R
+% Please edit documentation in R/vision-utils.R
 \name{coco_polygon_to_mask}
 \alias{coco_polygon_to_mask}
 \title{Convert COCO polygon to mask tensor (Robust Version)}
diff --git a/man/draw_bounding_boxes.Rd b/man/draw_bounding_boxes.Rd
index fbae4f1f..b2c3396e 100644
--- a/man/draw_bounding_boxes.Rd
+++ b/man/draw_bounding_boxes.Rd
@@ -1,5 +1,5 @@
 % Generated by roxygen2: do not edit by hand
-% Please edit documentation in R/vision_utils.R
+% Please edit documentation in R/vision-utils.R
 \name{draw_bounding_boxes}
 \alias{draw_bounding_boxes}
 \alias{draw_bounding_boxes.default}
@@ -21,6 +21,7 @@ draw_bounding_boxes(x, ...)
   width = 1,
   font = c("serif", "plain"),
   font_size = 10,
+  lazy = TRUE,
   ...
 )
 
@@ -34,6 +35,7 @@ draw_bounding_boxes(x, ...)
   width = 1,
   font = c("serif", "plain"),
   font_size = 10,
+  lazy = TRUE,
   ...
 )
 }
@@ -64,6 +66,9 @@ strings e.g. "red" or "#FF00FF". By default, viridis colors are generated for bo
 \item{font}{NULL for the current font family, or a character vector of length 2 for Hershey vector fonts.}
 
 \item{font_size}{The requested font size in points.}
+
+\item{lazy}{if \code{TRUE}, silently filter out degenerate bounding boxes (xmin >= xmax or ymin >= ymax).
+If \code{FALSE}, error on the first non-conforming box.}
 }
 \value{
 torch_tensor of size (C, H, W) of dtype uint8: Image Tensor with bounding boxes plotted.
@@ -78,7 +83,7 @@ image_tensor <- torch::torch_randint(170, 250, size = c(3, 360, 360))$to(torch::
 x <- torch::torch_randint(low = 1, high = 160, size = c(12,1))
 y <- torch::torch_randint(low = 1, high = 260, size = c(12,1))
 boxes <- torch::torch_cat(c(x, y, x + 20, y +  10), dim = 2)
-bboxed <- draw_bounding_boxes(image_tensor, boxes, colors = "black", fill = TRUE)
+bboxed <- draw_bounding_boxes(image_tensor, boxes, colors = "black", label = "label", fill = TRUE)
 tensor_image_browse(bboxed)
 }
 }
diff --git a/man/draw_keypoints.Rd b/man/draw_keypoints.Rd
index 03658131..d99683bd 100644
--- a/man/draw_keypoints.Rd
+++ b/man/draw_keypoints.Rd
@@ -1,5 +1,5 @@
 % Generated by roxygen2: do not edit by hand
-% Please edit documentation in R/vision_utils.R
+% Please edit documentation in R/vision-utils.R
 \name{draw_keypoints}
 \alias{draw_keypoints}
 \title{Draws Keypoints}
diff --git a/man/draw_segmentation_masks.Rd b/man/draw_segmentation_masks.Rd
index 0b6ccb12..9670398a 100644
--- a/man/draw_segmentation_masks.Rd
+++ b/man/draw_segmentation_masks.Rd
@@ -1,5 +1,5 @@
 % Generated by roxygen2: do not edit by hand
-% Please edit documentation in R/vision_utils.R
+% Please edit documentation in R/vision-utils.R
 \name{draw_segmentation_masks}
 \alias{draw_segmentation_masks}
 \alias{draw_segmentation_masks.default}
diff --git a/man/tensor_image_browse.Rd b/man/tensor_image_browse.Rd
index 3f226ea8..2bfef048 100644
--- a/man/tensor_image_browse.Rd
+++ b/man/tensor_image_browse.Rd
@@ -1,5 +1,5 @@
 % Generated by roxygen2: do not edit by hand
-% Please edit documentation in R/vision_utils.R
+% Please edit documentation in R/vision-utils.R
 \name{tensor_image_browse}
 \alias{tensor_image_browse}
 \title{Display image tensor}
diff --git a/man/tensor_image_display.Rd b/man/tensor_image_display.Rd
index 856cc3d0..98b24ddb 100644
--- a/man/tensor_image_display.Rd
+++ b/man/tensor_image_display.Rd
@@ -1,5 +1,5 @@
 % Generated by roxygen2: do not edit by hand
-% Please edit documentation in R/vision_utils.R
+% Please edit documentation in R/vision-utils.R
 \name{tensor_image_display}
 \alias{tensor_image_display}
 \title{Display image tensor}
diff --git a/man/vision_make_grid.Rd b/man/vision_make_grid.Rd
index 8fa91cf8..a113e20a 100644
--- a/man/vision_make_grid.Rd
+++ b/man/vision_make_grid.Rd
@@ -1,5 +1,5 @@
 % Generated by roxygen2: do not edit by hand
-% Please edit documentation in R/vision_utils.R
+% Please edit documentation in R/vision-utils.R
 \name{vision_make_grid}
 \alias{vision_make_grid}
 \alias{vision_make_grid.default}
diff --git a/tests/testthat/test-vision-utils.R b/tests/testthat/test-vision-utils.R
index 29d41e4f..aeb4d848 100644
--- a/tests/testthat/test-vision-utils.R
+++ b/tests/testthat/test-vision-utils.R
@@ -93,6 +93,49 @@ test_that("draw_bounding_boxes works", {
   expect_no_error(bboxed_image <- draw_bounding_boxes(image_uint, boxes, colors = "black", fill = TRUE))
 })
 
+test_that("draw_bounding_boxes lazy=TRUE silently filters degenerate boxes", {
+  image <- torch::torch_randint(1L, 200L, c(3L, 100L, 100L))$to(torch::torch_uint8())
+  valid_boxes <- torch::torch_tensor(rbind(c(10, 10, 50, 50), c(20, 20, 80, 80)))
+  bad_box     <- torch::torch_tensor(rbind(c(50, 10, 10, 50)))  # xmin > xmax
+  mixed_boxes <- torch::torch_cat(list(valid_boxes, bad_box), dim = 1)
+
+  expect_no_warning(result <- draw_bounding_boxes(image, mixed_boxes, lazy = TRUE))
+  expect_tensor_shape(result, c(3L, 100L, 100L))
+  expect_tensor_dtype(result, torch::torch_uint8())
+})
+
+test_that("draw_bounding_boxes lazy=FALSE errors on first degenerate box", {
+  image <- torch::torch_randint(1L, 200L, c(3L, 100L, 100L))$to(torch::torch_uint8())
+  # first box valid, second degenerate (index 2 should be reported)
+  boxes <- torch::torch_tensor(rbind(
+    c(10, 10, 50, 50),
+    c(50, 10, 10, 50)   # xmin > xmax
+  ))
+
+  expect_error(
+    draw_bounding_boxes(image, boxes, lazy = FALSE),
+    regexp = "xyxy format"
+  )
+  expect_error(
+    draw_bounding_boxes(image, boxes, lazy = FALSE),
+    regexp = "2"  # first invalid box index
+  )
+})
+
+test_that("draw_bounding_boxes lazy=TRUE warns when all boxes are degenerate", {
+  image <- torch::torch_randint(1L, 200L, c(3L, 100L, 100L))$to(torch::torch_uint8())
+  bad_boxes <- torch::torch_tensor(rbind(
+    c(50, 10, 10, 50),  # xmin > xmax
+    c(20, 80, 60, 20)   # ymin > ymax
+  ))
+
+  expect_warning(
+    result <- draw_bounding_boxes(image, bad_boxes, lazy = TRUE),
+    regexp = "No valid bounding box"
+  )
+  expect_tensor_shape(result, c(3L, 100L, 100L))
+})
+
 test_that("draw_bounding_boxes correctly mask a complete image", {
 
   image_float <- 1 - (torch::torch_randn(c(3, 360, 360)) / 20)
diff --git a/vignettes/examples/assets/dog_with_two_bbox.png b/vignettes/examples/assets/dog_with_two_bbox.png
deleted file mode 100644
index dab38255..00000000
Binary files a/vignettes/examples/assets/dog_with_two_bbox.png and /dev/null differ
diff --git a/vignettes/examples/assets/file84fb43ce0fa5.png b/vignettes/examples/assets/file84fb43ce0fa5.png
deleted file mode 100644
index 6ea0d76e..00000000
Binary files a/vignettes/examples/assets/file84fb43ce0fa5.png and /dev/null differ
diff --git a/vignettes/examples/assets/viz-boxes.jpg b/vignettes/examples/assets/viz-boxes.jpg
new file mode 100644
index 00000000..b602d787
Binary files /dev/null and b/vignettes/examples/assets/viz-boxes.jpg differ
diff --git a/vignettes/examples/assets/viz-fcn-classes.jpg b/vignettes/examples/assets/viz-fcn-classes.jpg
new file mode 100644
index 00000000..2971bee8
Binary files /dev/null and b/vignettes/examples/assets/viz-fcn-classes.jpg differ
diff --git a/vignettes/examples/assets/viz-fcn-dogs.jpg b/vignettes/examples/assets/viz-fcn-dogs.jpg
new file mode 100644
index 00000000..c5966fac
Binary files /dev/null and b/vignettes/examples/assets/viz-fcn-dogs.jpg differ
diff --git a/vignettes/examples/assets/viz-grid.jpg b/vignettes/examples/assets/viz-grid.jpg
new file mode 100644
index 00000000..837d2e4e
Binary files /dev/null and b/vignettes/examples/assets/viz-grid.jpg differ
diff --git a/vignettes/examples/assets/viz-keypoints.jpg b/vignettes/examples/assets/viz-keypoints.jpg
new file mode 100644
index 00000000..41cb7f50
Binary files /dev/null and b/vignettes/examples/assets/viz-keypoints.jpg differ
diff --git a/vignettes/examples/assets/viz-rfdetr-filtered.jpg b/vignettes/examples/assets/viz-rfdetr-filtered.jpg
new file mode 100644
index 00000000..6d1401eb
Binary files /dev/null and b/vignettes/examples/assets/viz-rfdetr-filtered.jpg differ
diff --git a/vignettes/examples/assets/viz-rfdetr-raw.jpg b/vignettes/examples/assets/viz-rfdetr-raw.jpg
new file mode 100644
index 00000000..121ec5f9
Binary files /dev/null and b/vignettes/examples/assets/viz-rfdetr-raw.jpg differ
diff --git a/vignettes/examples/visualization-utilities.Rmd b/vignettes/examples/visualization-utilities.Rmd
new file mode 100644
index 00000000..0b3ae0da
--- /dev/null
+++ b/vignettes/examples/visualization-utilities.Rmd
@@ -0,0 +1,197 @@
+---
+title: "Visualization utilities"
+type: docs
+---
+
+```{r, echo = FALSE}
+knitr::opts_chunk$set(eval = FALSE)
+```
+
+This article walks through the drawing helpers shipped with torchvision
+(`vision_make_grid()`, `draw_bounding_boxes()`, `draw_segmentation_masks()` and
+`draw_keypoints()`) and shows how to feed them the output of a pretrained
+detection or segmentation model. It follows the
+[Visualization utilities](https://docs.pytorch.org/vision/stable/auto_examples/others/plot_visualization_utils.html)
+tutorial from PyTorch.
+
+The code below is not evaluated when the website is built, because it downloads
+around 250 MB of model weights. Every figure was produced by running it.
+
+```{r setup}
+library(torchvision)
+library(torch)
+
+gallery <- "https://raw.githubusercontent.com/pytorch/vision/main/gallery/assets/"
+dog1 <- base_loader(paste0(gallery, "dog1.jpg")) %>% transform_to_tensor()
+dog2 <- base_loader(paste0(gallery, "dog2.jpg")) %>% transform_to_tensor()
+```
+
+## A grid of images
+
+`vision_make_grid()` tiles a batch of images into a single tensor, which is a
+quick way to eyeball a batch coming out of a `dataloader()`.
+
+```{r}
+grid <- vision_make_grid(dog1, dog2, per_row = 2)
+tensor_image_browse(grid)
+```
+
+![The two photographs tiled side by side into a single tensor](assets/viz-grid.jpg)
+
+## Bounding boxes
+
+`draw_bounding_boxes()` expects an `(N, 4)` tensor of boxes given as
+`(xmin, ymin, xmax, ymax)` in absolute pixel coordinates.
+
+```{r}
+boxes <- torch_tensor(rbind(
+  c(50, 50, 100, 200),
+  c(210, 150, 350, 430)
+), dtype = torch_float())
+
+boxed <- draw_bounding_boxes(dog1, boxes, 
+                             labels = c("tree", "partial corgi"), font_size = 18, 
+                             colors = c("darkorange", "yellow"), width = 5)
+tensor_image_browse(boxed)
+```
+
+![Hand-written boxes: orange over the trees, yellow around the corgi](assets/viz-boxes.jpg)
+
+## Object detection with RF-DETR
+
+Boxes are usually not hand written but predicted. We will use the RF-DETR model to do so.
+RF-DETR expects ImageNet-normalized inputs, so we assemble two images in a batch and then normalize.  
+Note: The model is able to detect the dogs on unnormalized input, but with far lower confidence: the corgi confidence scores in the previous image drops from 0.67 with normalization down 0.28 without.
+
+```{r}
+norm_mean <- c(0.485, 0.456, 0.406)
+norm_std <- c(0.229, 0.224, 0.225)
+
+batch <- list(dog1, dog2) %>%
+  torch_stack() %>% 
+  transform_normalize(norm_mean, norm_std)
+
+model <- model_rfdetr_base(pretrained = TRUE, num_queries = 40)
+model$eval()
+detections <- with_no_grad(model(batch))$detections
+```
+
+RF-DETR resizes the batch to its own working resolution internally and maps the
+boxes back to the coordinates of the images you passed in, so images can be drawn
+straight onto the original tensors. Each element of `detections` holds `boxes`,
+`labels` and `scores` sorted by decreasing score. 
+
+Let's first draw the objects detected with raw model output. 
+
+```{r}
+tensor_image_browse(vision_make_grid(
+  draw_bounding_boxes(dog1, detections[[1]]$boxes),
+  draw_bounding_boxes(dog2, detections[[2]]$boxes)
+))
+```
+![RF-DETR raw output has many boxes](assets/viz-rfdetr-raw.jpg)
+
+We can see that too many boxes are detected. RF-DETR Model default to 300 object detection, but we reduced it via the `num_queries` parameter. 
+
+Let's filter them based on their associated confidence score.
+
+We will create a convenience function that keeps the confident ones, and
+turn the label indices into readable names out of `coco_classes()`.
+
+```{r}
+draw_high_confidence_bboxes <- function(image, detection, width = 4, font_size = 25, ...) {
+  keep <- as.logical(as.array(detection$scores > 0.2))
+  draw_bounding_boxes(
+    image,
+    detection$boxes[keep, , drop = FALSE],
+    labels = coco_classes(as.integer(detection$labels[keep])), 
+    width = width,
+    font_size = font_size,
+    ...
+  )
+}
+```
+
+Now let's draw again
+```{r}
+tensor_image_browse(vision_make_grid(
+  draw_high_confidence_bboxes(dog1, detections[[1]], color = "yellow"),
+  draw_high_confidence_bboxes(dog2, detections[[2]], color = "yellow")
+))
+
+```
+
+  ![RF-DETR confidence filtered output](assets/viz-rfdetr-filtered.jpg)
+
+## Semantic segmentation masks
+
+`model_fcn_resnet50()` scores every pixel against the 21 PASCAL VOC classes. Its
+pretrained weights are evaluated at 520 pixels, so resize before normalizing,
+and keep the resized images around to draw on: masks and image must share their
+height and width.
+
+```{r}
+resized <- list(dog1, dog2) %>%
+  torch_stack() %>%
+  transform_resize(size = c(520, 520)) 
+
+batch <- resized %>%
+  transform_normalize(norm_mean, norm_std)
+
+model <- model_fcn_resnet50(pretrained = TRUE)
+model$eval()
+scores <- with_no_grad(model(batch))$out
+```
+
+`scores` is a `CPUFloatType{2,21,520,520}` tensor. Handeling such a tensor,
+`draw_segmentation_masks()` takes the `argmax()` over the class dimension and
+colors each class that appears.
+
+```{r}
+segmented <- draw_segmentation_masks(resized[1,..], scores[1, ..], alpha = 0.6)
+tensor_image_browse(segmented)
+```
+
+![Every class the model assigns: the dog in yellow, the background in pink](assets/viz-fcn-classes.jpg)
+
+To highlight a single class instead, build the boolean mask yourself. `dog` is
+the 13th of the 21 classes, and softmax turns the scores into per-pixel
+probabilities that can be thresholded.
+
+```{r}
+tensor_image_browse(vision_make_grid(
+  draw_segmentation_masks(resized[1,..], scores[1, ..], alpha = 0.6),
+  draw_segmentation_masks(resized[2,..], scores[2, ..], alpha = 0.6)
+  )
+)
+```
+
+![The dog class on its own](assets/viz-fcn-dogs.jpg)
+
+## Keypoints
+
+`draw_keypoints()` takes an `(N, K, 2)` tensor holding the `(x, y)` location of
+`K` keypoints for each of `N` instances. torchvision has no pose estimation
+model to feed it (`model_mtcnn()` predicts five landmarks, but only on human
+faces), so the keypoints here are annotated by hand: the two ear tips, the two
+eyes and the nose of the corgi. `connectivity` lists the pairs of keypoint
+indices to join with a line. Each line is drawn in the color of the keypoint it
+starts from, so leave `colors` at its default or pass one color per keypoint.
+
+```{r}
+keypoints <- torch_tensor(rbind(
+  c(220, 171),
+  c(296, 176),
+  c(239, 222),
+  c(262, 226),
+  c(242, 250)
+), dtype = torch_float())$reshape(c(1, 5, 2))
+
+connectivity <- list(c(1, 3), c(2, 4), c(3, 4), c(3, 5), c(4, 5))
+
+posed <- draw_keypoints(dog1, keypoints, connectivity = connectivity,
+                        radius = 6, width = 3)
+tensor_image_browse(posed)
+```
+
+![Five hand-annotated keypoints joined by the connectivity pairs](assets/viz-keypoints.jpg)