FitAQA

A Benchmark of Fitness Action Quality Assessment for Multimodal Large Language Models

FitAQA is a systematic benchmark for evaluating multimodal large language models (MLLMs) on fitness action quality assessment. It tests whether a model can recognize quality-relevant visual evidence, judge exercise execution, and localize form errors over time.

This repository contains annotations and metadata only. Videos are not redistributed. Users must obtain media from the original providers and comply with each source dataset's terms.

Dataset Overview

Item	Count
Referenced videos	2,219
Video files with SHA-256	2,163
Bodyweight exercises	30
QA instances	5,512
Perception instances	2,562
Judgement instances	2,562
Temporal-grounding instances	388
Unified form-error types	38
Quality dimensions	6

Source composition

Subset	Source	QA rows	Unique videos
Short clips	QEVD-FIT-300k	4,192	1,551
Short clips	EgoExo-Fitness	418	183
Short clips	Kinetics-700	402	186
Short clips	UCF101	112	56
Long videos	QEVD-FIT-COACH	115	83
Long videos	QEVD-FIT-COACH-Benchmark	62	44
Long videos	QEVD-FIT-COACH-Competition	211	116

Benchmark Tasks

1. Perception asks which observable manifestation of a target quality aspect is present in a short clip. Options are neutral visual descriptions.
2. Judgement asks whether the same quality aspect is executed correctly. Every judgement item is paired with a perception item through the numeric suffix of question_id.
3. Temporal grounding asks a model to locate every interval in a longer single-exercise clip where a described form error occurs.

Configurations and Schema

multiple_choice (default)

The test split contains all 5,124 perception and judgement instances.

Field	Type	Description
question_id	string	Unique ID: perception_NNNNNN or judgement_NNNNNN.
task	string	perception or judgement.
dataset	string	Original video source.
video_name	string	Source asset name or frame-sequence identifier.
video_sha256	string or null	SHA-256 of the referenced video file; null for directory-backed frame sequences.
action_name	string	Target exercise.
form_error	list[string]	Form-error label(s) assessed by the question pair.
question	string	Question shown to the model.
options	list[{label, text}]	Options in their fixed presentation order.
gold	string	Correct option label.
correct_form_option	string	Option describing correct execution; not necessarily gold.

The stored option order is part of the benchmark. Do not reshuffle options at evaluation time.

temporal_grounding

The test split contains 388 instances over 243 processed single-exercise videos and 526 annotated temporal segments.

Field	Type	Description
question_id	string	Unique grounding_NNNNNN ID.
task	string	Always temporal_grounding.
dataset	string	One of the three QEVD-FIT-COACH sources.
video_name	string	Processed single-exercise clip identifier.
video_sha256	string	SHA-256 of the processed single-exercise video file.
action_name	string	Exercise performed in the clip.
form_error	list[string]	Unified form-error label.
error_detail	string	Video-specific description of the target error.
gold	list[{start, end}]	All ground-truth intervals in seconds.

Loading the Dataset

from datasets import load_dataset

mcq = load_dataset(
    "Kelly0510/FitAQA",
    "multiple_choice",
    split="test",
)

grounding = load_dataset(
    "Kelly0510/FitAQA",
    "temporal_grounding",
    split="test",
)

The files can also be read directly with PyArrow, pandas, Polars, or DuckDB.

Video Sources and Preprocessing

Videos are deliberately excluded from this release. Each file-backed row includes video_sha256, computed from the corresponding file under <video-root>/<dataset>/<video_name>. Directory-backed frame sequences are not hashed and use null.

Source	How video_name is resolved
QEVD-FIT-300k	Same filename as the source dataset.
EgoExo-Fitness	Same filename as the source dataset.
Kinetics-700	Same filename as the source dataset.
UCF101	Same source identifier; retained samples are represented as 1 fps frame-sequence directories.
QEVD-FIT-COACH, Benchmark, Competition	Processed clip name produced by the procedure below.

The three long-video sources contain multi-exercise workout sessions. FitAQA creates candidate single-exercise clips from the official dense feedback, transition annotations, and per-frame timestamp arrays:

1. Collapse repeated non-empty frame-aligned feedback strings into temporal spans.
2. Use the first frame of the current transition feedback as the inclusive clip start.
3. Use the midpoint frame of the next transition feedback as the inclusive clip end.
4. Keep candidate intervals containing at least one non-transition feedback.
5. Number valid candidates within each source workout as {source_stem}_{index:02d}.mp4.
6. Recover the effective frame rate as number_of_timestamps / (last_timestamp - first_timestamp) and encode H.264 MP4 clips.

The standalone script in this repository implements that procedure without project-specific paths:

python preprocess.py \
  --input-root /path/to/QEVD \
  --output-root /path/to/output \
  --splits QEVD-FIT-COACH \
           QEVD-FIT-COACH-Benchmark \
           QEVD-FIT-COACH-Competition

Use --dry-run to generate segments.jsonl manifests without invoking FFmpeg, and --overwrite to replace existing clips. The script requires NumPy and a system installation of FFmpeg.

This preprocessing produces candidate action clips. FitAQA subsequently retained clips whose execution quality varies over time and subjected them to human annotation and expert review. The 243 clips referenced by dataset and video_name in the temporal_grounding configuration are the final benchmark subset.

Unified Form-Error Taxonomy

The taxonomy below is a human-readable reference for interpreting labels in the benchmark. It is documentation, not a separate dataset configuration.

ID	Dimension	Form Error	Example Manifestation
1	Alignment	Excessive Ankle Eversion	Medial arch collapse during vertical jump landing.
2	Alignment	Insufficient Ankle Dorsiflexion	Heels lifted off the floor in squat or downward-facing dog.
3	Alignment	Insufficient Knee Flexion	Knees stay too straight in quick feet.
4	Alignment	Excessive Knee Flexion	Working leg bent in side-lying leg raise.
5	Alignment	Excessive Hip Internal Rotation	Knees collapse inward during weight-bearing lower-body movements.
6	Alignment	Excessive Hip External Rotation	Knees turned too far outward in squat or glute bridge.
7	Alignment	Excessive Hip Abduction	Knee flares too far outward during leg drive in mountain climbers.
8	Alignment	Excessive Hip Adduction	Knee crosses too far inward during leg drive in mountain climbers.
9	Alignment	Insufficient Hip Flexion	Torso too upright from limited hip hinge in quick feet.
10	Alignment	Excessive Hip Flexion	Torso leans too far forward in forward lunge, side lunge, or jumping lunge.
11	Alignment	Insufficient Shoulder External Rotation	Elbows rotate inward during arm circles or standing YTW.
12	Alignment	Insufficient Shoulder Internal Rotation	Elbow points forward instead of outward in overhead triceps stretch.
13	Alignment	Insufficient Shoulder Flexion	Shoulder angle remains too closed in downward-facing dog support.
14	Alignment	Excessive Shoulder Abduction	Elbows flared too wide from the torso in push-up.
15	Alignment	Excessive Scapular Elevation	Shoulders shrugged toward the ears in plank or downward-facing dog.
16	Alignment	Insufficient Scapular Protraction	Shoulder blades not spread enough in plank or push-up.
17	Alignment	Excessive Elbow Flexion	Elbows too bent in posterior deltoid stretch or downward dog support.
18	Alignment	Excessive Lumbar Flexion	Low back rounds in squat or push-up.
19	Alignment	Excessive Lumbar Extension	Low back sags in push-up, bird dog extension, or reverse crunch lowering.
20	Alignment	Excessive Thoracic Flexion	Rounded upper back in plank, push-up, or downward-facing dog.
21	Alignment	Excessive Cervical Flexion	Head lifted with chin tucked in dead bug or reverse crunch.
22	Alignment	Excessive Cervical Extension	Gaze lifted toward the ceiling instead of forward in squat.
23	Alignment	Forward Head Posture	Head reaches forward toward the floor in push-up.
24	Alignment	Incorrect Movement Trajectory	Arms drift forward out of the side plane during standing YTW transitions.
25	Alignment	Incorrect Support Placement	Improper hand spacing in push-up or stance width in squat.
26	Symmetry	Load Distribution Asymmetry	Uneven shoulder height during downward-facing dog support.
27	Symmetry	Range-of-Motion Asymmetry	Uneven leg height in high knees, or asymmetric limb lowering in dead bug.
28	Stability	Trunk Instability	Trunk sways side to side with alternating legs in mountain climbers.
29	Stability	Support Instability	Unstable landing support in jumping lunge, or shaky leg in glute bridge.
30	Coordination	Movement Incoordination	Unsynchronized arm-leg motion in jumping jack, or same-side limbs used in dead bug.
31	Tempo	Loss of Eccentric Tempo Control	Letting the leg and torso drop during alternating leg V-up descent.
32	Tempo	Momentum-Driven Execution	Swinging the legs to complete reverse crunch.
33	Tempo	Low Movement Cadence	Slow leg lift or long pause between steps in high knees.
34	Completeness	Insufficient Range of Motion	Limited arm raise, arm return, or leg spread in jumping jack.
35	Completeness	Missing Movement Phase	Missing push-up or jump phase in burpee.
36	Completeness	Insufficient Movement Duration	No hold after entering downward-facing dog.
37	Completeness	Extraneous Movement	Extra heel lift during glute bridge ascent, or head lift during descent.
38	Completeness	Exercise Mismatch	Performing substantially different from the requested exercise.

Evaluation

• Perception: exact option-label accuracy and question-macro accuracy.
• Judgement: recall, precision, and F1 with erroneous execution as the positive class.
• Temporal grounding: mean temporal IoU and Recall at IoU thresholds 0.3, 0.5, and 0.7, using the benchmark's multi-interval matching protocol.

Present the stored options directly. For temporal grounding, predict every interval where the described manifestation occurs.

License

The FitAQA annotations, Parquet files, documentation, and independently implemented preprocessing script in this repository are released under the Apache License 2.0.

No source videos are included or relicensed. All rights in third-party media remain with their respective owners and providers.

Acknowledgements

FitAQA builds on videos from QEVD, EgoExo-Fitness, Kinetics-700, and UCF101. The public Qualcomm FitCoach loader was consulted to understand the QEVD annotation layout; preprocess.py is an independent minimal implementation of the FitAQA preprocessing procedure.

Citation

Please cite the FitAQA paper when using this benchmark:

@article{zheng2026fitaqa,
  title   = {FitAQA: A Benchmark of Fitness Action Quality Assessment for Multimodal Large Language Models},
  author  = {Zheng, Kaili and Wang, Kaiwen and Zhu, Xun and Yang, Qingyuan and Guo, Chenyi and Wu, Ji},
  journal = {arXiv preprint arXiv:2608.08736},
  year    = {2026},
  url     = {https://arxiv.org/abs/2608.08736}
}

Paper: FitAQA: A Benchmark of Fitness Action Quality Assessment for Multimodal Large Language Models