about the keypoint model

[NeKoooo233]

First congratulations on the release of the yolo26-pose model. I have seen many keypoint model projects, and your project is one of the best, from annotation to training, from single-objective to multi-objective. Not just for the human body, but can be extended to key point detection of any object.
I've come to the conclusion that keypoint detection and localization may be the cornerstone of other keypoint tasks, such as face detection and pose estimation.
In my opinion, many face detection methods still haven't broken free from the framework of prior estimation like SIFT. Given a completely black photo, some face models can still make predictions, which is unacceptable.
In real-world applications, pixel-level accuracy is still required, such as in surgical procedures and industrial object measurement. However, many current models still use mAP as their evaluation method. My question is whether using average pixel error to evaluate the model would be better. But, this would require a pixel-level labeled dataset.
In short, if the community's research direction shifts towards keypoint localization, I think there might be even more interesting discoveries, just as instance segmentation shifted towards segmenting anything.

[UltralyticsAssistant]

👋 Hello @NeKoooo233, thank you for sharing your thoughtful perspective on keypoint detection and localization, and for your interest in the YOLO26-pose model 🚀! We recommend a visit to the Docs for new users where you can find many Python and CLI usage examples and where many of the most common questions may already be answered.

If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us debug it.

If this is a custom training ❓ Question or research discussion, please provide as much information as possible, including dataset image examples, keypoint annotation details, evaluation requirements, and training logs, and verify you are following our Tips for Best Training Results.

Join the Ultralytics community where it suits you best. For real-time chat, head to Discord 🎧. Prefer in-depth discussions? Check out Discourse. Or dive into threads on our Subreddit to share knowledge with the community.

This is an automated response 🤖, and an Ultralytics engineer will also assist you soon.

Upgrade

Upgrade to the latest ultralytics package including all requirements in a Python>=3.8 environment with PyTorch>=1.8 to verify your issue is not already resolved in the latest version:

pip install -U ultralytics

Environments

YOLO may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all Ultralytics CI tests are currently passing. CI tests verify correct operation of all YOLO Modes and Tasks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

[NeKoooo233]

Some of the materials I found
https://dtic.dimensions.ai/details/publication/pub.1200712744

[NeKoooo233]

This may mean that the larger the model, the more it relies on prior knowledge during reasoning.

[glenn-jocher]

Not necessarily—model size alone does not prove greater reliance on learned priors; compare variants under identical blank, corrupted, and out-of-distribution inputs using false-positive rate and confidence calibration. For precision-critical tasks, report normalized pixel error alongside the standard OKS-based pose mAP described in the COCO-Pose documentation, since raw pixel error varies with image scale; physical measurement additionally requires camera calibration.