• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Wednesday, November 18, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Health

Pose Geometry Fails Independent Predictive Evaluation on Sports Performance Quality Benchmark

by
November 12, 2026
in Biology, Health
Reading Time: 4 mins read
0
Pose Geometry Fails Independent Predictive Evaluation on Sports Performance Quality Benchmark

Pose Geometry Fails Independent Predictive Evaluation on Sports Performance Quality Benchmark

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A methodological audit published in Frontiers in Physiology reveals that transparent pose-derived image-plane geometry does not provide reproducible predictive value for expert-rated athletic performance when evaluated on an independent test split. Researchers Yi Wang, Ke He, Yongchao Li, and Lei Shi examined whether interpretable geometric features extracted from video poses could reliably score sports actions once model development, hyperparameter selection, and final testing were strictly separated. Although action quality assessment has increasingly relied on high-capacity video architectures, such black-box models frequently obscure the precise visual cues driving automated scores. The authors sought to establish whether a fully transparent geometric alternative could reproduce judging outcomes under rigorous leakage controls.

To test this question, the research team analyzed the established AQA-7 sports benchmark, which encompasses 1,189 video recordings across seven athletic disciplines and a total of 165,212 decoded image frames. The study preserved the official Split_4 configuration of the benchmark, which designated 803 videos for development and 303 videos for independent final testing, while excluding 83 trampoline recordings that fell outside the supervised partition. Every frame was processed through the MediaPipe Pose Landmarker version 1.0.1 running on central processing units in dedicated video tracking mode. Across the full collection of frames, only 48,344 individual frames, representing 29.3 percent of the total, yielded at least one detected human pose, highlighting the substantial technical friction in deriving continuous skeletal landmarks from dynamic broadcast video footage.

The study then applied a sequence of outcome-independent pose and geometry quality-control stages that severely constrained the usable sample volume. From the initial 803 development videos, preliminary quality screening retained 131 samples suitable for geometric feature construction. Imposing a fixed completeness threshold requiring 57 predefined geometric features further winnowed the development pool down to 108 viable samples. For the final inferential development phase, the dataset was constrained to 102 samples representing only two sports: 48 competitive diving actions and 54 gymnastic vault routines. These remaining samples were subjected to repeated nested cross-validation to tune model parameters while isolating them from the final evaluation data.

The experimental framework compared three distinct modeling paradigms: an action-only baseline, a geometry-only Ridge regression model, and a combined geometry-plus-action Ridge regression model. To eliminate data leakage and optimistic bias, the researchers locked all development configurations, feature extraction rules, and model weights before processing the final test split. Score-blind test-set features and model predictions were generated and formally registered prior to unblinding the official final-test performance scores. In the independent test set, only 33 diving and vault samples satisfied all pose completeness criteria and were evaluable for final out-of-sample predictive assessment.

During the internal development phase, the geometry-only Ridge regression demonstrated a modest rank-order relationship with expert judging scores. Development validation yielded a macro Spearman rank correlation coefficient of 0.231 with a 95 percent confidence interval spanning from 0.039 to 0.417, alongside a Pearson correlation coefficient of 0.227 with a 95 percent confidence interval from 0.020 to 0.409. Despite these positive correlation metrics, the development model produced a negative coefficient of determination of -0.574, signaling that the raw geometric predictions were already poorly calibrated to the scale and dispersion of expert evaluations.

When the frozen geometry-only Ridge model was applied to the independent final test split, its apparent ordinal performance failed to replicate entirely. On the 33 evaluable test trials, the model produced an inverted Spearman rank correlation of -0.116 with a wide 95 percent confidence interval ranging from -0.499 to 0.375, and a negligible Pearson correlation of 0.042 with a confidence interval between -0.355 to 0.404. Predictive error exploded on the unseen data, generating a coefficient of determination of -11.986, a standardized mean absolute error of 1.322, and a root mean squared error of 1.778. In stark contrast, the simple action-only comparator achieved markedly superior test performance, recording a coefficient of determination of -0.076, a standardized mean absolute error of 0.478, and a root mean squared error of 0.597.

Further diagnostics indicated that geometric feature validity was heavily sport dependent, registering a Cramer’s V of 0.412 across the development data, reflecting major differences in landmark detection fidelity between diving and vaulting kinematics. However, this pose-tracking validity did not reflect athletic quality: the action-adjusted difference in expert scores between valid and invalid video samples was only 0.047 standard deviations, with a 95 percent confidence interval of -0.148 to 0.233 and a permutation p-value of 0.874. The researchers concluded that frozen two-dimensional image-plane geometry cannot deliver out-of-sample predictive value beyond basic action category baselines, underscoring that the tested pipeline serves as a critical audit of pose estimation limitations rather than a deployable automated scoring platform.

Subject of Research: Evaluation of pose-derived image-plane geometry for automated sports action quality assessment

Article Title: Pose-derived image-plane geometry for expert-rated sports performance quality: leakage-controlled development and independent evaluation on AQA-7

Article References: Wang, C., Feng, X., Wang, Y., He, K., Li, Y., Shi, L., & Liang, S. (2026). Pose-derived image-plane geometry for expert-rated sports performance quality: leakage-controlled development and independent evaluation on AQA-7. Frontiers in Physiology, 17, Article 1873326. https://doi.org/10.3389/fphys.2026.1873326

Image Credits: AI Generated

DOI: 10.3389/fphys.2026.1873326

Keywords: action quality assessment, AQA-7 benchmark, pose estimation, MediaPipe Pose Landmarker, image-plane geometry, Ridge regression, sports performance evaluation, computer vision in sports, model leakage control, independent validation, diving assessment, gymnastics vault scoring, Spearman rank correlation, Pearson correlation, standardized mean absolute error, root mean squared error, cross-validation, methodological audit

News Source: Drew Townsend. (November 12, 2026). Pose Geometry Fails Independent Predictive Evaluation on Sports Performance Quality Benchmark. Scienmag.

Tags: action quality assessmentAQA-7 benchmarkcomputer vision in sportscross-validationdiving assessmentgymnastics vault scoringimage-plane geometryindependent validationMediaPipe Pose Landmarkermethodological auditmodel leakage controlPearson correlationpose estimationridge regressionroot mean squared errorSpearman rank correlationsports performance evaluationstandardized mean absolute error
Share12Tweet7Share2ShareShareShare1

Related Posts

Gallic Acid Shows Multi-Target Potential Against Breast Cancer in Comprehensive Review

Gallic Acid Shows Multi-Target Potential Against Breast Cancer in Comprehensive Review

November 18, 2026
Herbal Formulation Reduces Brain Bleeding and Gut Damage in Rat Stroke Model

Herbal Formulation Reduces Brain Bleeding and Gut Damage in Rat Stroke Model

November 18, 2026

Lawsonia Intracellularis Infection Triggers Rare Granulomatous Ileitis in Growing Farm Pig

November 18, 2026

Scapular Depression Is Independent of Rockwood Grade in ACJ Injuries

November 18, 2026

POPULAR NEWS

  • Epilepsy Marks Lower Adaptive Functioning in Boys With MECP2 Duplication Syndrome

    Epilepsy Marks Lower Adaptive Functioning in Boys With MECP2 Duplication Syndrome

    29 shares
    Share 12 Tweet 7
  • Geno2pheno Overestimates CXCR4 Tropism in Chinese HIV-1 CRF01_AE Clinical Isolates

    29 shares
    Share 12 Tweet 7
  • Climate Variability Drives Growing Global Risks Across Foodborne Pathogen Pathways

    29 shares
    Share 12 Tweet 7
  • Catheter Drainage Reduces Surgical Burden While Endoscopy Accelerates Early Hemorrhage Clearance

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Gallic Acid Shows Multi-Target Potential Against Breast Cancer in Comprehensive Review

Disabling Cellular Recycling Boosts Production of Hard-to-Manufacture Membrane Proteins

Herbal Formulation Reduces Brain Bleeding and Gut Damage in Rat Stroke Model

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 88 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.