diff --git a/torchvision/models/detection/roi_heads.py b/torchvision/models/detection/roi_heads.py
index 51b210cb6f3..cbca6b8364b 100644
--- a/torchvision/models/detection/roi_heads.py
+++ b/torchvision/models/detection/roi_heads.py
@@ -232,9 +232,47 @@ def _onnx_heatmaps_to_keypoints_loop(
 
 
 def heatmaps_to_keypoints(maps, rois):
-    """Extract predicted keypoint locations from heatmaps. Output has shape
-    (#rois, 4, #keypoints) with the 4 rows corresponding to (x, y, logit, prob)
-    for each keypoint.
+    """Extract predicted keypoint locations from heatmaps.
+
+    This function takes as input heatmaps and regions of interest (ROIs) and extracts
+    predicted keypoint locations. The output has two tensors:
+    
+    - The first tensor has the shape (#rois, #keypoints, 3), where each row corresponds
+      to a keypoint and contains the following information:
+      - x: The x-coordinate of the keypoint location in the image.
+      - y: The y-coordinate of the keypoint location in the image.
+      - vis: The visibility score of the keypoint (usually 1 for visible keypoints).
+
+    - The second tensor has shape (#rois, #keypoints), where each value corresponds to
+      the logit score of a keypoint. Logits are raw scores that represent the confidence
+      of the model in predicting the keypoint's location.
+
+    Args:
+        maps (torch.Tensor): A tensor containing heatmaps for keypoints. It has the shape
+            (#rois, #keypoints, height, width).
+        rois (torch.Tensor): A tensor containing regions of interest (ROIs) for which
+            keypoints are predicted. Each row in `rois` represents an ROI and should
+            have the format (x1, y1, x2, y2), where (x1, y1) are the coordinates of the
+            top-left corner and (x2, y2) are the coordinates of the bottom-right corner.
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: A tuple containing two tensors:
+        - xy_preds (torch.Tensor): The predicted keypoint locations with shape
+            (#rois, #keypoints, 3), where the last dimension corresponds to (x, y, vis).
+        - end_scores (torch.Tensor): The confidence scores (logits) for each predicted
+            keypoint with shape (#rois, #keypoints).
+
+    The `xy_preds` tensor contains the estimated keypoint locations in the image coordinate
+    system, while the `end_scores` tensor provides confidence scores for each keypoint.
+    You can use the `xy_preds` to obtain the actual keypoint locations in the image, and
+    the `end_scores` to assess the model's confidence in its predictions.
+
+    Note:
+    - The `vis` value in `xy_preds` typically represents the visibility of the keypoint,
+      where 1 indicates a visible keypoint and 0 indicates an occluded or invisible
+      keypoint.
+    - You can threshold the `end_scores` to filter out keypoints with low confidence.
+
     """
     # This function converts a discrete image coordinate in a HEATMAP_SIZE x
     # HEATMAP_SIZE image to a continuous keypoint coordinate. We maintain