Add fasterrcnn documentation

_pkgdown.yml

@@ -29,8 +29,10 @@ navbar:
         href: articles/examples/texture-nca.html
       - text: fcnresnet
         href: articles/examples/fcnresnet.html
-      - text: keypoints
-        href: articles/examples/keypoints.html
+      - text: fasterrcnn
+        href: articles/examples/fasterrcnn.html
+      - text: keypoint
+        href: articles/examples/keypoint.html
 
 reference:
 - title: Transforms

vignettes/examples/fasterrcnn.R

@@ -0,0 +1,138 @@
+# Overview ---------------------------------------------------------
+#
+# model_fasterrcnn_resnet50_fpn() implements the Faster R-CNN
+# object detection architecture with a ResNet-50 backbone and
+# Feature Pyramid Network (FPN), combining region proposal,
+# classification, and bounding box regression in a single pipeline.
+
+
+# Loading Images ---------------------------------------------------
+library(torchvision)
+library(torch)
+
+url1 <- "https://raw.githubusercontent.com/pytorch/vision/main/gallery/assets/dog1.jpg"
+url2 <- "https://raw.githubusercontent.com/pytorch/vision/main/gallery/assets/dog2.jpg"
+
+dog1 <- magick_loader(url1) %>% transform_to_tensor()
+dog2 <- magick_loader(url2) %>% transform_to_tensor()
+
+
+# Visualizing a grid of images -------------------------------------
+
+
+dogs <- torch_stack(list(dog1, dog2))
+grid <- vision_make_grid(dogs, scale = TRUE, num_rows = 2)
+tensor_image_browse(grid)
+
+
+# Preprocessing the data -------------------------------------
+
+
+norm_mean <- c(0.485, 0.456, 0.406)
+norm_std  <- c(0.229, 0.224, 0.225)
+
+dog1_prep <- dog1 %>%
+  transform_resize(c(520,520)) %>%
+  transform_normalize(mean = norm_mean, std = norm_std)
+dog2_prep <- dog2 %>%
+  transform_resize(c(520,520)) %>%
+  transform_normalize(mean = norm_mean, std = norm_std)
+
+# make batch (2,3,520,520)
+dog_batch <- torch_stack(list(dog1_prep, dog2_prep))
+
+
+# Loading Model -------------------------------------
+
+# pretrained = TRUE loads weights trained on the COCO dataset
+# (Common Objects in Context benchmark for object detection)
+# score_thresh filters detections with confidence below the threshold
+model <- model_fasterrcnn_resnet50_fpn(pretrained = TRUE, score_thresh = 0.5)
+model$eval()
+
+# run model
+output <- model(dog_batch)
+
+
+# Understanding the Output Structure -------------------------------
+
+# In eval() mode, output contains:
+names(output)  # typically "detections" in eval() mode
+
+# output$detections: list with one element per image in batch
+
+detections <- output$detections
+length(detections)  # equals the batch size (2 images)
+
+# Each detection contains boxes, labels, and scores
+det1 <- detections[[1]]
+names(det1)  # "boxes", "labels", "scores"
+
+# boxes: (N, 4) tensor with [x1, y1, x2, y2] pixel coordinates
+# representing top-left and bottom-right corners
+det1$boxes$shape
+
+# labels: (N) tensor with COCO class IDs
+det1$labels$shape
+
+# scores: (N) tensor with confidence values [0, 1]
+det1$scores$shape
+
+# Note: In train() mode, output contains losses instead:
+# model$train()
+# output <- model(images, targets)
+# output$losses  # list(loss_classifier, loss_box_reg, loss_objectness, loss_rpn_box_reg)
+
+
+# Visualizing the Output ------------------------------
+
+
+num_boxes1 <- as.integer(det1$boxes$size()[1])
+if (num_boxes1 > 0) {
+  keep <- seq_len(min(5, num_boxes1))
+  boxes1 <- det1$boxes[keep, ]
+  # coco_label() converts COCO class IDs to readable class names
+  labels1 <- coco_label(as.integer(det1$labels[keep]))
+  scores1 <- as.numeric(det1$scores[keep])
+  labels_with_scores1 <- paste0(labels1, " (", sprintf("%.2f", scores1), ")")
+
+  detected1 <- draw_bounding_boxes(
+    dog1 %>% transform_resize(c(520,520)),
+    boxes = boxes1,
+    labels = labels_with_scores1,
+    width = 3
+  )
+
+  tensor_image_browse(detected1)
+}
+
+det2 <- detections[[2]]
+num_boxes2 <- as.integer(det2$boxes$size()[1])
+if (num_boxes2 > 0) {
+  keep <- seq_len(min(5, num_boxes2))
+  boxes2 <- det2$boxes[keep, ]
+  labels2 <- coco_label(as.integer(det2$labels[keep]))
+  scores2 <- as.numeric(det2$scores[keep])
+  labels_with_scores2 <- paste0(labels2, " (", sprintf("%.2f", scores2), ")")
+
+  detected2 <- draw_bounding_boxes(
+    dog2 %>% transform_resize(c(520,520)),
+    boxes = boxes2,
+    labels = labels_with_scores2,
+    width = 3
+  )
+
+  tensor_image_browse(detected2)
+}
+
+
+# Summary ----------------------------------------------------------
+#
+# This article demonstrated:
+# - Loading and preprocessing images for object detection
+# - Running Faster R-CNN inference with pretrained weights
+# - Understanding detection output structure (boxes, labels, scores)
+# - Visualizing bounding boxes with confidence scores and class labels
+#
+# Faster R-CNN provides high-quality instance-level object detection
+# suitable for both inference and training workflows.

vignettes/examples/fasterrcnn.Rmd

@@ -0,0 +1,9 @@
+---
+title: "fasterrcnn"
+type: docs
+---
+
+```{r, echo = FALSE}
+knitr::opts_chunk$set(eval = FALSE)
+knitr::spin_child(paste0(rmarkdown::metadata$title, ".R"))
+```