Diving-QA: Quantitative Diving Assessment

• Diving-QA is the integrated research program that applies quantitative video analysis, action quality assessment, and scoring models to competitive diving.
• It leverages methodologies like low-cost barycenter tracking, temporal segmentation, and uncertainty-aware score prediction to improve performance evaluation.
• Advanced neural and neuro-symbolic frameworks are used to enhance score accuracy, interpretability, and fairness by addressing judge subjectivity and biomechanical factors.

Diving-QA can be understood as the body of research that evaluates competitive diving through quantitative video analysis, action quality assessment (AQA), and scoring models. In this literature, the central problem is not only fine-grained dive recognition, but also the estimation of execution quality under temporal structure, judge subjectivity, and fairness constraints. Recent work therefore spans low-cost barycenter tracking, temporal localization and dive-code classification, procedure-aware and uncertainty-aware score prediction, neuro-symbolic explanation, and biomechanical or statistical analyses of fairness in judging (Frassinelli et al., 2017, Xu et al., 2022, Zhou et al., 2022, Okamoto et al., 2024).

1. Problem formulation and scope

Competitive diving presents an unusually demanding AQA setting because the action class is fine-grained, the diver often occupies a very small fraction of the frame, and the scoring process is partly subjective. Earlier AQA formulations emphasized direct score regression from video, but later work argued that this is less reliable because it obscures the internal temporal structure of a dive and provides poor interpretability (Xu et al., 2022). A related line of work observed that most AQA methods predict a single deterministic score even though real competitions involve multiple judge scores and therefore measurable diversity in expert assessment (Zhou et al., 2022).

The field consequently divides into several tightly connected subproblems. One concerns upstream perception: extracting individual dive clips from continuous footage, tracking the diver, and classifying the dive code (Nibali et al., 2017). A second concerns quantitative performance representation: barycenter trajectories, maximum height, entry position, and temporally segmented sub-actions (Frassinelli et al., 2017, Frassinelli et al., 2017). A third concerns score modeling proper: regressing final scores, predicting score differences relative to exemplars, or generating distributions over plausible judge scores rather than a single point estimate (Parmar et al., 2019, Xu et al., 2022, Zhou et al., 2022). A fourth concerns fairness: whether difficulty coefficients, body morphology, or judge behavior introduce systematic discrepancies that should be measured or corrected (Sun, 2018, McGee, 2023).

This suggests that Diving-QA is best viewed not as a single model family but as an integrated research program combining perception, representation, inference, and evaluation.

2. Data resources and annotation regimes

Dataset construction has been central because diving quality assessment requires labels that are richer than a single final score. The literature progresses from score-only datasets toward multi-task, step-level, and pose-level supervision.

Dataset	Scale and coverage	Annotation focus
MTL-AQA	1412 samples from 16 different international competitions	Official action quality score, factorized action class, expert commentary
FineDiving	3000 video samples from 30 events; 52 action types, 29 sub-action types, 23 difficulty ratings	Step-level temporal annotation, official action scores, individual judges’ scores, degree of difficulty
FineDiving-Pose	12,722 manually annotated frames in 367 videos; 288,000 automatically annotated frames	16 keypoints for improved diving pose labels
Diving48	18,404 video clips and 48 classes	Fine-grained action classification benchmark

The MTL-AQA dataset was introduced to support joint learning of score regression, fine-grained action recognition, and commentary generation, under the hypothesis that richer supervision improves representations for quality assessment (Parmar et al., 2019). FineDiving extended this logic by introducing a two-level semantic structure and a two-level temporal structure. It provides action-level annotation, precise step boundaries, official action scores, individual judges’ scores, and degree of difficulty, all sourced from FINA records (Xu et al., 2022). FineDiving-Pose further addressed the observation that existing pose labels for diving were low quality because of motion blur, occlusion, and confusion with spectators; it combined manual annotation with automatic annotation to improve skeletal supervision (Qi et al., 7 Jan 2025).

These datasets encode different assumptions about what constitutes relevant evidence. MTL-AQA emphasizes multi-task semantics and commentary; FineDiving emphasizes procedure structure and judge-level scoring; FineDiving-Pose emphasizes body configuration; Diving48 isolates fine-grained recognition without direct quality labels (Kanojia et al., 2019). A plausible implication is that benchmark choice strongly conditions which aspect of diving performance a model can learn to explain.

3. Classical video pipelines: sampling, tracking, and clip extraction

Before the recent AQA models, competitive diving analysis was framed as a low-cost video-processing problem. One line of work proposed a pipeline of four main stages: video sampling, frame projection on panorama, barycenter identification, and barycenter trajectory tracking (Frassinelli et al., 2017). A related paper expressed the full process as “Video Sampling > Image Mosaicking > Barycenter Identification > Barycenter Trajectory > Performance Score” and emphasized flexibility and low cost, including operation on standard commercial PCs and off-the-shelf video (Frassinelli et al., 2017).

Image mosaicking compensates for camera vibration and moving viewpoints by aligning frames into a common reference background. Barycenter identification then uses HSV color filtering, subtraction of the filtered panorama from the filtered frame, and centroid computation of the detected diver region. The resulting temporal sequence is smoothed with a moving average filter. From this representation, the literature extracts barycenter trajectory, barycenter at water entry, and maximum barycenter height as quantitative performance cues (Frassinelli et al., 2017). The approach was designed to work for both 10 m platform and 3 m springboard diving, with typical acquisition rates such as 20 fps or 25 Hz depending on the formulation (Frassinelli et al., 2017, Frassinelli et al., 2017).

A more complete automation pipeline addressed continuous video footage. It used a volumetric 3D CNN, TALNN, on a 21-frame sliding window to predict three temporal signals—start, middle, and end—followed by smoothing with a Hann window and interval extraction. Reported performance improved from a segment-based baseline with precision 76.7%, recall 71.6%, and $F_1$ 74.1% to precision 88.3%, recall 98.3%, and $F_1$ 92.9% at IoU $\geq 0.5$, with 98% of dives correctly extracted (Nibali et al., 2017). The same work treated diver localization as tracking-by-detection and found that a segmentation-based approach outperformed direct regression; it then imposed global physical constraints through linear and quadratic trajectory fitting with RANSAC. Dive-code prediction used a multi-head 3D CNN so that rotation type, pose type, somersault number, twist number, and handstand start were predicted as separate outputs rather than as one sparse combined class (Nibali et al., 2017).

These classical pipelines remain important because they formalize the perception subproblems that later AQA systems inherit, even when the downstream model is fully learned.

4. Neural action quality assessment architectures

The transition from classical pipelines to neural AQA was marked by a shift from coarse video-level regression toward structured, multi-task, and stage-aware learning. One influential formulation treated diving AQA as a multitask problem with three outputs: action quality score regression, fine-grained action recognition, and commentary generation. On MTL-AQA, the C3D-AVG-MTL approach achieved a rank correlation of 90.44%, outperforming its single-task counterpart and earlier baselines (Parmar et al., 2019). The core claim was that representations learned only for action recognition are not sufficient for AQA and instead should be learned jointly with quality-relevant supervision.

FineDiving pushed this further by explicitly modeling procedure structure. Its Temporal Segmentation Attention module first parses query and exemplar videos into consecutive steps, then applies procedure-aware cross-attention between corresponding steps, and finally performs fine-grained contrastive regression. On FineDiving, TSA reported [email protected] of 80.71 and [email protected] of 30.17 without dive number matching, alongside Spearman’s $\rho = 0.8925$ and $R$-$\ell_2 = 0.4782$; with dive number matching, the corresponding figures were [email protected] of 82.51, [email protected] of 34.31, $\rho = 0.9203$, and $R$-$\ell_2 = 0.3420$ (Xu et al., 2022). The paper’s central methodological claim was that understanding both high-level semantics and internal temporal structures of actions is key to accurate and interpretable predictions.

UD-AQA addressed a different limitation: the subjectivity of judging. It introduced a CVAE-based module conditioned on the video and judge score, used latent-space sampling to produce multiple plausible judge scores, and re-weighted the regression loss by estimated uncertainty. On FineDiving, UD-AQA reported Spearman correlation 0.9341 and $R$-$F_1$0 (×100) 0.2985, improving over TSA-15 at 0.9204 and 0.3419, respectively (Zhou et al., 2022). Its uncertainty-guided training strategy further ordered training from low-uncertainty to high-uncertainty samples.

More recent work combined procedure segmentation with pose guidance. The hierarchical pose-guided multi-stage contrastive regression framework introduced a multi-scale dynamic visual-skeleton encoder, a Bi-Directional GRU-based procedure segmentation network, a multi-modal fusion module, and stage-wise contrastive regression. On FineDiving without dive numbers it reported SRCC 0.9041, $F_1$1 (×100) 0.411, and [email protected] / @0.75 of 98.17 / 93.68; with dive numbers the reported values were SRCC 0.9365, $F_1$2 0.244, and [email protected] / 0.75 of 99.46 / 97.18 (Qi et al., 7 Jan 2025). The work specifically argued that fixed frame splits disrupt the temporal continuity of sub-actions and lead to unavoidable prediction errors.

Across these models, the dominant trend is clear: performance gains are associated with stronger inductive structure, whether through multi-task supervision, procedure-aware step alignment, uncertainty modeling, or explicit skeleton-guided stage decomposition.

5. Explainability and neuro-symbolic assessment

Interpretability has become a central concern because purely neural AQA systems are trained on subjective human judgements and often provide limited transparency. FineDiving already emphasized this by making step-wise segmentation and cross-attention maps part of the scoring mechanism; the model’s visualizations indicate which spatial and temporal regions of the exemplar are aligned with the query in each procedure step (Xu et al., 2022). This provides a structured explanation of why a predicted score differs from a reference score.

A more explicit response was the hierarchical neuro-symbolic approach, NS-AQA. It decomposes the system into a neural action-context parser and a rules-based action analyzer. The parser extracts interpretable symbols such as platform location, splash mask, and 2D pose; derived symbols include joint angles and distance between diver and platform. The analyzer then applies domain-expert-defined micro-programs to recognize the dive, segment it into phases, and assess fine-grained quality elements such as distance from platform, feet apart, somersault tightness, over/under-rotation, and splash size (Okamoto et al., 2024).

NS-AQA also defines an objective scoring methodology. Each quality element is converted to a percentile relative to a reference population, and the overall score is computed by uniform averaging,

$F_1$3

The system automatically generates a detailed visio-linguistic HTML report with per-element explanations and visual evidence. In expert evaluation, six national/international level judges, coaches, and divers reviewed 50 dives each; experts agreed with the system on all categories $F_1$4 of the time and preferred NS-AQA output to neural models in 96.1% of cases (Okamoto et al., 2024).

The explainability literature therefore does not merely add post hoc visualization. It redefines the assessment target itself, from a single scalar output toward a decomposed account of what happened in takeoff, flight, and entry, and why each element influenced the final judgment.

6. Fairness, biomechanics, and judging discrepancies

Diving-QA also includes research on whether existing scoring conventions are fair. One biomechanical analysis argued that FINA’s traditional difficulty coefficient reflects only the complexity of the maneuver and does not fully consider athlete body shape. It modeled the human body as six rigid parts—trunk, head, two arms, and two legs—used cuboid approximations and the parallel axis theorem to compute moments of inertia, and derived completion times for pike, tuck, and twist under constant angular momentum (Sun, 2018). The key result was that completion time is proportional to $F_1$5, which was proposed as the body shape correction index.

Using female-athlete reference points, the paper derived a body shape correction coefficient through Lagrange interpolation,

$F_1$6

and proposed an adjusted scoring rule,

$F_1$7

In this scheme, slim, small athletes with low $F_1$8 have $F_1$9, larger athletes with high $\geq 0.5$0 have $\geq 0.5$1, and the average case remains at $\geq 0.5$2 (Sun, 2018). The paper presented this as a new 10-meter platform diving difficulty coefficient scheme.

A different fairness question concerns judges rather than morphology. An analysis of high-school one meter diving competitions in the US from 2017 to 2022 defined the diver competency score as the mean net score over all dives and meets,

$\geq 0.5$3

and discrepancy as

$\geq 0.5$4

The dataset contained 39,105 dives, 261 meets, and 1911 divers (McGee, 2023). The reported findings were that median discrepancies for both genders aligned at zero, discrepancy increased slightly with age, tuck position dives were judged more leniently, and—contrary to expectations—easier dives were judged more leniently while difficult dives were judged more harshly. Round effects were also systematic: as the meet progressed from round 1 to 11, discrepancy decreased by about 0.08–0.09 per round (McGee, 2023).

Taken together, these two lines of work show that “difficulty” is not a single construct. It can refer to rotational biomechanics, embodied morphology, or the way judging practice interacts with degree of difficulty and competition order.

7. Limitations and research directions

The literature identifies several persistent limitations. Classical video pipelines remain limited by frame extraction frequency, color segmentation quality, the lack of direct 3D kinematics, and vulnerability to occlusion or similarity between diver and background (Frassinelli et al., 2017, Frassinelli et al., 2017). Upstream recognition is also intrinsically difficult because the diver may occupy fewer than 1% of the pixels in each frame, while other divers and bystanders can still be present (Nibali et al., 2017).

Modern AQA systems inherit different limitations. End-to-end neural models can lack transparency and may encode biases because they are trained on subjective human judgements (Okamoto et al., 2024). Fixed frame segmentation can disrupt temporal continuity of sub-actions, which motivated procedure segmentation and variable-length stage modeling (Qi et al., 7 Jan 2025). Pose estimation errors remain a failure mode for explainable systems, especially under water or with motion blur (Okamoto et al., 2024). Judge diversity further implies that a single deterministic target score can be an incomplete supervisory signal, which motivated uncertainty-driven modeling (Zhou et al., 2022).

A plausible implication is that future Diving-QA systems will continue to converge on hybrid formulations: procedure-aware temporal structure from FineDiving-style models, multimodal pose and appearance encoding, explicit uncertainty over judge scores, and rule-based decompositions that expose the contribution of each dive element. In the current literature, the most durable pattern is that improved performance and improved interpretability have increasingly been pursued together rather than as competing objectives.