IJCV 2026

A Decade of Action Quality Assessment:
Largest Systematic Survey of Trends, Challenges, and Future Directions

Hao Yin1,2*  ·  Paritosh Parmar3*  ·  Daoliang Xu1,2  ·  Yang Zhang2
Tianyou Zheng2 ✉  ·  Weiwei Fu1,2 ✉
1 USTC   2 SIBET, CAS   3 IHPC, A*STAR, Singapore
* Equal contribution ✉ Corresponding authors
📰 IJCV Paper 💻 GitHub Repo 📝 BibTeX

§ Abstract

Action Quality Assessment (AQA)—the ability to quantify the quality of human motion, actions, or skill levels and provide feedback—has far-reaching implications in areas such as low-cost physiotherapy, sports training, and workforce development. As such, it has become a critical field in computer vision and video understanding over the past decade. In this paper, we present a thorough survey of the AQA landscape, systematically reviewing over 200 research papers using the PRISMA framework. We begin by covering foundational concepts and definitions, then move to general frameworks and performance metrics, and finally discuss the latest advances in methodologies and datasets. This survey provides a detailed analysis of research trends, performance comparisons across 33 datasets and 7 principal research trends, challenges, and future directions. Through this work, we aim to offer a valuable resource for both newcomers and experienced researchers, promoting further exploration and progress in AQA.

Action Quality Assessment Skills Assessment Video Understanding Computer Vision Deep Learning Systematic Survey PRISMA

★ Survey at a Glance

214
Papers Reviewed
33
Public Datasets
9
Application Domains
7
Principal Trends
2014–2025
Coverage Period

📊 Open Overview Figure (PDF)
Taxonomy overview. The AQA landscape: problem definition, datasets across 9 domains, 7 methodology trends, and future directions.

📊 Open AQA Pipeline (PDF)
AQA pipeline. General framework: from input modalities through feature extraction, quality assessment, and downstream feedback.

1 AQA Datasets — 33 Benchmarks Across 9 Domains

Following PRISMA guidelines, we identify 33 publicly available datasets spanning 9 domains since 2014. The field has grown from 2 datasets in 2014 to 33 today — a 16.5× expansion in a decade. Click 🔗 for dataset or project links.
DatasetYearSamplesAvg. Dur.AnnotationDomainLink
MIT-Dive2014159~2.5sScore Sport 🔗
MIT-Skate2014150~175sScore Sport 🔗
JIGSAWS2014103~92sScore, Action Surgery 🔗
UNLV-Dive2017370~3.8sScore Sport 🔗
UNLV-Vault2017176~2.8sScore Sport 🔗
UI-PRMD2018100—Grade, Action Rehab
EPIC-Skill2018216~86.6sRank, Action Daily 🔗
AQA-720191,189~6.7sScore, Action Sport 🔗
MTL-AQA20191,412~4.1sScore, Action, Desc. Sport 🔗
Fis-V2019500~170sScore Sport 🔗
BEST2019500~187.6sGrade, Rank, Action Daily 🔗
KIMORE2019353~29.9sScore, Action Rehab
TASD-22020606~4.1sScore, Action Sport
Rhythmic Gym.20201,000~95sScore, Action Sport 🔗
PISA (Piano-Skills)2021992~160frGrade, Difficulty Music 🔗
FR-FS2021417~103frGrade, Action Sport 🔗
FS100020211,000—Score, Action Sport 🔗
SMART20215,000~420frScore, Action Sport
SimSurgSkill2021315—Score, Action Surgery
Fitness-AQA202221,284~4.1sGrade, Action Fitness 🔗
FineDiving20223,000~4.2sScore, Action, Step Sport 🔗
Assembly10120224,321~426sScore, Action Industrial 🔗
LOGO2023200~204.2sScore, Action, Form. Sport 🔗
FineFS20231,167~215sScore, Action Sport 🔗
PaSk20231,018~10.7sScore Sport
CDRG2023240~14.7sRank, Action Dance
GAIA20249,180~2.8sScore, Action AIGV 🔗
EgoExo-4D20245,035~312sGrade, Action Daily 🔗
EgoExo-Learn20243,304~10sRank, Action Daily 🔗
EgoExo-Fitness20246,131~18.8sScore, Action, Desc. Fitness 🔗
AVOS20241,997—Grade, Action Surgery
UJ-AQA20258,540~28frScore Sport 🔗
BASKET202532,232~500sGrade, Action Sport 🔗
FLEX20257,512~234frScore, Action, Desc., sEMG Fitness 🔗

📊 View Sample Figure (PDF)
Action samples across 9 dataset domains.

📊 View Sankey Chart (PDF)
Sankey diagram. Dataset publication flow by year × domain.

Domain Distribution

17
Sport
3
Surgery
2
Rehab
4
Daily
1
Music
3
Fitness
1
Industrial
1
Dance
1
AIGV

3 Challenges & Future Directions

Despite dramatic progress over the past decade, significant challenges remain. At the action level, inherent complexity and the lack of a true "gold standard" reference limit generalization. At the dataset level, the field needs larger scale, broader subject/action diversity, physiological multimodality (e.g., sEMG, IMU), and richer annotation granularity. At the methodology level, key open problems include real-time inference, unified evaluation metrics, interpretability, multimodal robustness under missing modalities, and ultimately moving toward General AQA Agents capable of multi-domain assessment with safety-governance for high-stakes applications.

§ Citation

If you find this survey useful, please cite our IJCV 2026 paper:

@article{yin2026decadeactionqualityassessment,
  title     = {A Decade of Action Quality Assessment: Largest Systematic Survey of Trends, Challenges, and Future Directions},
  author    = {Yin, Hao and Parmar, Paritosh and Xu, Daoliang and Zhang, Yang and Zheng, Tianyou and Fu, Weiwei},
  journal   = {International Journal of Computer Vision (IJCV)},
  year      = {2026},
  doi       = {10.1007/s11263-025-02672-4},
  url       = {https://github.com/HaoYin116/Survey_of_AQA}
}
Copied to clipboard!