TTA: Occlusion-Rich Table Tennis Dataset

• TTA is an occlusion-rich table tennis tracking dataset featuring manually annotated 2D ball coordinates from professional Paralympic matches.
• It employs a detailed visibility taxonomy—fully visible, partially occluded, and fully occluded—enabling robust performance evaluation.
• The dataset benchmarks temporal inference methods with metrics like RMSE and accuracy, highlighting performance under challenging occlusion scenarios.

Searching arXiv for the TTA dataset and closely related table-tennis datasets to ground the article in the cited literature. Searching arXiv for "Table Tennis Australia TOTNet TTA occlusion-rich dataset". Searching arXiv for related datasets: OpenTTGames, TTStroke-21, and P²ANet for context and comparison. The Table Tennis Australia dataset, abbreviated TTA, is an occlusion-rich table tennis tracking dataset introduced as a benchmark for ball tracking under partial and full occlusions in realistic play. It was developed in collaboration with Paralympics Australia, collected from professional-level Paralympic table tennis matches, and positioned as a resource for evaluating single-view ball localization under challenging visibility conditions. In the paper that introduces it, TTA is defined by 9,159 samples at 25 fps and $1080 \times 1920$ resolution, with 1,996 occlusion samples, and serves as the central table-tennis benchmark for the occlusion-aware tracking model TOTNet (Xu et al., 13 Aug 2025).

1. Identity and provenance

TTA is explicitly presented as “a new occlusion-rich table tennis dataset” and as “a manually annotated table tennis dataset” collected from four professional-level Paralympic table tennis matches. The work was developed “in collaboration with Paralympics Australia”, and the labels were reviewed with national team analysts to ensure quality (Xu et al., 13 Aug 2025).

Its motivation is tightly coupled to a specific failure mode of racket-sport perception systems: existing datasets often contain limited occlusion cases, lower frame rates, or no explicit visibility labels. TTA was therefore designed to reflect real-world gameplay conditions in which the ball is frequently hidden by players, equipment, or viewpoint geometry. The paper presents this as a dataset for object tracking performance under challenging occlusion scenarios, rather than as a stroke-recognition, rally-outcome, or multimodal biomechanics corpus (Xu et al., 13 Aug 2025).

A recurrent source of confusion in the literature is that several table-tennis resources are sometimes informally grouped under generic abbreviations. In strict bibliographic terms, however, the TTA name is used in the TOTNet work; other table-tennis papers describe different datasets under different names, including OpenTTGames, TTStroke-21, and P$^2$ANet (Voeikov et al., 2020, Martin et al., 2021, Bian et al., 2022).

2. Acquisition protocol and dataset structure

The dataset is reported as comprising 9,159 samples at 25 fps and $1080 \times 1920$ resolution, drawn from four professional-level Paralympic table tennis matches. The benchmark is explicitly described as occlusion-rich, with 1,996 occlusion samples, attributed to camera angles, ball size, and dynamic play (Xu et al., 13 Aug 2025).

The annotation protocol follows standard visibility labels, with three categories used in the TTA benchmark:

• Fully Visible
• Partially Occluded
• Fully Occluded

The paper provides the following split-level visibility breakdown for TTA (Xu et al., 13 Aug 2025):

Split	Fully Visible	Partially Occluded	Fully Occluded
Train	5,141	834	650
Validation	1,285	215	158
Test	668	72	67

In the reported experiments, the tracking model uses a configurable number of input frames, with five frames selected as the best trade-off between temporal context and efficiency. A plausible implication is that TTA is intended not merely as a framewise ball-localization set, but as a benchmark for short-horizon temporal inference under occlusion (Xu et al., 13 Aug 2025).

The paper does not provide a match-by-match metadata inventory, nor does it enumerate player identities, venues, or camera calibration parameters. It does state, however, that the data come from professional-level Paralympic matches and were curated specifically to benchmark tracking in real-world conditions rather than in laboratory or minimally occluded settings (Xu et al., 13 Aug 2025).

3. Annotation scheme and benchmark formulation

TTA provides 2D ball coordinates and visibility labels for each annotated target frame. The visibility taxonomy distinguishes fully visible, partially occluded, and fully occluded cases, enabling both stratified evaluation and visibility-aware training (Xu et al., 13 Aug 2025).

In the TOTNet benchmark formulation, ground truth is represented differently depending on visibility. For fully visible and partially occluded frames, the target is encoded as one-hot heatmaps over the horizontal and vertical image axes. For fully occluded frames, the target becomes a normalized Gaussian distribution centered at the labeled position, explicitly modeling positional uncertainty under total visual absence (Xu et al., 13 Aug 2025).

This label structure is used in a visibility-weighted BCE loss: $$L = w_v \cdot \left( \mathrm{BCE}(P_x, T_{x,\mathrm{map}}) + \mathrm{BCE}(P_y, T_{y,\mathrm{map}}) \right),$$ where $v$ indexes visibility level and $w_v$ is the associated weight. The dataset therefore supports training regimes in which occluded frames contribute differently from visible ones, rather than being treated as ordinary localization errors (Xu et al., 13 Aug 2025).

Evaluation on TTA uses two principal metrics:

• RMSE, computed from Euclidean pixel error between predicted and labeled ball coordinates.
• Accuracy, defined by a distance threshold:
  • 5 pixels for fully visible and partially occluded frames,
  • 10 pixels for fully occluded frames.

This asymmetric thresholding is a consequential design choice. It encodes the fact that full occlusion is not simply a harder version of ordinary detection, but a distinct inference regime in which label uncertainty and temporal reasoning are central (Xu et al., 13 Aug 2025).

4. Empirical role in the TOTNet benchmark

TTA is the central table-tennis benchmark used to compare TTNet, TrackNetV2, monoTrack, WASB, and TOTNet under visibility-stratified evaluation. Its main contribution is not merely aggregate difficulty, but the fact that it reveals large performance separations specifically in the fully occluded regime (Xu et al., 13 Aug 2025).

On TTA, the strongest reported model is TOTNet (OF), which combines temporal modeling with optical flow. Its reported performance is:

• Visible: RMSE 1.84, Accuracy 0.98
• Partially Occluded: RMSE 3.45, Accuracy 0.92
• Fully Occluded: RMSE 7.19, Accuracy 0.80

By contrast, the strong baseline WASB reports:

• Visible: RMSE 2.14, Accuracy 0.97
• Partially Occluded: RMSE 5.26, Accuracy 0.88
• Fully Occluded: RMSE 37.30, Accuracy 0.63

The fully occluded case is the decisive one: TTA exposes a large gap between methods that remain competent when the ball disappears entirely and methods that primarily interpolate visible detections (Xu et al., 13 Aug 2025).

The ablation study is also conducted on TTA. It shows that visibility-weighted BCE improves full-occlusion performance, occlusion augmentation becomes effective only when combined with such weighting, and adding optical flow further reduces fully occluded RMSE from 12.31 to 7.19 while increasing accuracy from 0.74 to 0.80 (Xu et al., 13 Aug 2025). This indicates that TTA is not merely a passive evaluation set; it is the empirical substrate on which the paper’s modeling claims are established.

5. Relation to other table-tennis datasets

TTA occupies a distinct niche within the table-tennis dataset landscape. It is neither a stroke-taxonomy benchmark nor a multimodal biomechanics corpus; it is a visibility-annotated ball-tracking benchmark centered on occlusion robustness (Xu et al., 13 Aug 2025).

Several other datasets address different problems:

OpenTTGames provides 120 fps $1920 \times 1080$ videos with event labels, semantic segmentation masks, and ball coordinates for multi-task analysis, but the TTA paper characterizes the TTNet/OpenTTGames-style setting as having minimal occlusion and lacking explicit visibility labels (Voeikov et al., 2020, Xu et al., 13 Aug 2025).

TTStroke-21 is a video benchmark for fine-grained stroke detection and classification, with 20 stroke classes in the MediaEval formulation and HD 120 fps recordings under natural conditions. Its annotation target is stroke boundaries and stroke identity, not ball visibility under occlusion (Martin et al., 2021, Martin et al., 2023).

P$^2$ANet is a broadcast-based dense action detection benchmark with 2,721 video clips, 139,075 labeled stroke segments, and 14 fine-grained action classes. It addresses dense temporal action localization and recognition from 25 FPS broadcasting videos rather than ball localization under visibility uncertainty (Bian et al., 2022).

RacketVision extends the problem space toward ball tracking, racket pose estimation, and trajectory forecasting, and its table-tennis subset contains 50 games, 780 clips, 170,027 frames, 19,495 ball annotations, and 6,648 racket annotations. It broadens multimodal sports analytics, but its principal novelty is ball–racket interaction modeling rather than occlusion-stratified ball tracking (Dong et al., 21 Nov 2025).

A common misconception is therefore to treat “table tennis dataset” as a single category. TTA is much narrower and more specialized: it is a dataset for occlusion-aware 2D ball tracking, particularly valuable when the research question is whether a model can infer ball location when direct visual evidence is partial or absent (Xu et al., 13 Aug 2025).

6. Access, constraints, and research significance

The paper states that “The dataset will be shared upon academic request” and points to the project repository at https://github.com/AugustRushG/TOTNet for code and access information. No explicit public license is specified in the paper text (Xu et al., 13 Aug 2025).

This access model matters because TTA is built from professional-level Paralympic matches and incorporates annotations reviewed with national team analysts. The paper does not formalize a long-term benchmark server, leaderboard, or redistribution policy. As a result, TTA currently appears closer to a controlled academic resource than to an unrestricted public archive (Xu et al., 13 Aug 2025).

Its research significance lies in three properties. First, it provides a visibility-aware annotation protocol that cleanly separates fully visible, partially occluded, and fully occluded cases. Second, it supplies enough difficult samples—especially 1,996 occlusion samples—to make fully occluded tracking a measurable problem rather than an anecdotal edge case. Third, it shifts table-tennis ball tracking away from pure detection and toward temporal inference under uncertainty, which is precisely the regime needed for downstream applications such as post-match analytics, coaching, and referee support (Xu et al., 13 Aug 2025).

In that sense, TTA is best understood not as a general-purpose table-tennis dataset, but as a specialized benchmark that operationalizes one of the field’s hardest perceptual conditions: recovering the ball state when the ball is small, fast, and frequently hidden.