Does labelling the whole object help in better detection of the parts?

[Shravasti221]

I take an example from another user's project where they detect parts of the face. Would this model perform better if the face also was annotated as a separate label? Or would the overlapping bounding boxes confide the model?

project credits: https://github.com/ignaciohrdz/yolo-face-parts-detector

I ask this as I am working on a project that detects body and head separately, and I am not able to get good results for the detection of the head label. So I was wondering if I marked the entire character also (3 labels: body, head, character) would it perform with higher recall? I am currently training with images at 1920px resolution.

The green is body, yellow is head, pink is human

[UltralyticsAssistant]

👋 Hello @Shravasti221, thank you for your question! Adding a full-body or full-character label may provide additional context, but it does not guarantee better head recall. For the best results, annotate every head consistently and ensure the head boxes are accurate, sufficiently varied in size and appearance, and well represented across the dataset. A useful approach is to compare training with and without the additional character class while keeping the validation set unchanged. Also verify the inference confidence threshold and image size, since small head targets can be sensitive to both.

Please share a few representative images and your training logs or configuration if you would like more specific guidance, and verify that you are following the Tips for Best Training Results. This is an automated response, and an Ultralytics engineer will assist you soon.

[glenn-jocher]

Overlapping nested boxes for different classes are valid and should not inherently confuse YOLO, but adding a character box does not explicitly teach the model that its head is inside it, so it is unlikely to improve head recall by itself—add that class only if you need to detect it. Focus instead on complete, tight, consistent head annotations, sufficient and varied head examples, class balance, and the actual training imgsz (1920px source images are resized to imgsz, so consider a larger imgsz or tiling/cropping if heads become small); see the annotation best practices.

[rishig47-dev]

If the goal is to detect the individual face parts, I would not add a face class just to improve the detection of 'eyes', 'nose', or 'mouth'.

YOLO can handle overlapping bounding boxes, so the parts can be annotated independently. Adding a whole-face class is useful only if you also need the model to detect the complete face as a separate object.

For better part detection, I would focus more on consistent annotations, enough examples for each class, class balance, and sufficient image resolution—especially because eyes and other facial parts are relatively small objects.

So, if your required outputs are only the facial parts, I would start with separate classes such as 'eyes', 'nose', or 'mouth rather than adding an overlapping face class.

[pderrenger]

Adding a character box is unlikely to improve head recall by itself; use it only if you need character detections. Overlapping parent/part boxes are valid, but prioritize labeling every head consistently with tight boxes and verify its pixel size after resizing to the training imgsz—higher imgsz, crops/tiling, more representative head examples, or pose keypoints may help more for small heads. See the annotation best practices.