CourtSI-Ext: Cross-Sport Spatial Benchmark

• CourtSI-Ext is a cross-sport evaluation benchmark defined to test the transfer of spatial reasoning from traditional net sports to pickleball.
• It reuses the CourtSI taxonomy and a 3D reconstruction pipeline to ensure consistent evaluation metrics across different court sports.
• Empirical results show significant gains in overall accuracy with fine-tuning, though challenges in localization persist.

CourtSI-Ext is a cross-sport evaluation benchmark for spatial intelligence in vision-LLMs (VLMs), introduced within the CourtSI framework as an “unseen-sport evaluation set” built on pickleball rather than the training sports of badminton, tennis, and table tennis. It reuses the same taxonomy, annotation conventions, and evaluation protocol as CourtSI-Bench, so that measured performance reflects transfer across sport domains rather than changes in task format. In the reported release, CourtSI-Ext contains 215 QA pairs, 111 images, and 35 distinct scenes, and is used to test whether court-geometry-grounded spatial reasoning learned from CourtSI transfers to a similar but previously unseen net sport (Yang et al., 10 Mar 2026).

1. Position within the CourtSI framework

CourtSI-Ext is defined relative to two other datasets in the same benchmark family. CourtSI is the large-scale training corpus; CourtSI-Bench is the in-domain evaluation set over the same seen sports; CourtSI-Ext is the out-of-domain evaluation benchmark over pickleball, which is treated as an unseen but structurally similar net sport (Yang et al., 10 Mar 2026).

Dataset	QA pairs	Images / scenes
CourtSI	1,008,941	52,481 / 1,057
CourtSI-Bench	3,686	1,988 / 382
CourtSI-Ext	215	111 / 35

This placement gives CourtSI-Ext a specific methodological role. It is not presented as a new training corpus, and it does not introduce a separate reasoning taxonomy. Instead, it functions as a controlled transfer benchmark: the question schema is held fixed while the sport domain changes. The paper does not report scene-overlap specifics for CourtSI-Ext, but it explicitly treats the benchmark as disjoint in sport domain from CourtSI training data. A common misunderstanding is to view CourtSI-Ext as merely a smaller version of CourtSI-Bench; the paper instead positions it as a diagnostic split for cross-sport transfer.

2. Data sources and construction pipeline

CourtSI-Ext is produced using the same semi-automatic 3D reconstruction pipeline used for CourtSI. The benchmark is built from YouTube videos, and the supplementary details state that the videos include both men’s and women’s matches, as well as singles and doubles, in order to keep the image distribution relatively balanced. All videos are 1080p at 25 fps, matching the source-data characteristics used in CourtSI (Yang et al., 10 Mar 2026).

The reconstruction process inherits the same geometry-based assumptions and calibration procedure. The paper states that net sports provide fixed court dimensions, and that court corners and net points serve as metric anchors. Camera parameters are estimated via PnP from manually annotated 2D court keypoints and known 3D court coordinates. The resulting coordinate system is court-anchored and metric. Ball positions are reconstructed by converting image clicks and ground projections into 3D via ray-plane intersection, while player meshes are recovered with PromptHMR and then depth-corrected by annotating the lowest vertex height. CourtSI-Ext applies this same engine to pickleball broadcast imagery rather than redesigning the pipeline for a different sport.

This reuse of the data engine is central to the benchmark’s interpretation. Because CourtSI-Ext is constructed under the same reconstruction logic, differences between CourtSI-Bench and CourtSI-Ext are intended to expose failures of transfer rather than mismatches in annotation workflow or metric grounding. The paper’s qualitative material, including the figure labeled “Court-Ext examples,” serves to illustrate that the benchmark is based on real broadcast pickleball footage with court geometry similar to tennis and badminton.

3. Taxonomy and geometric scope

CourtSI-Ext directly inherits the CourtSI-Bench taxonomy. The benchmark covers four top-level task families: distance measurement, spatial counting, localization, and relational reasoning (Yang et al., 10 Mar 2026).

Within these families, the same subtypes are reused. The paper lists distance measurement subtypes such as camera-object, object-object, object-line, and height; counting subtypes such as ball counting and player counting; localization as object localization; and relational categories including ball-zone, ball-player, camera-player, player-zone, player-player, and player-line. No new taxonomy is defined for CourtSI-Ext. The only major shift is the sport domain.

The relevance of court geometry is methodological rather than merely descriptive. The paper emphasizes that CourtSI questions are grounded in world coordinates made available by court-anchored reconstruction. For CourtSI-Ext, pickleball is suitable precisely because its court structure remains sufficiently similar to the seen net sports for the original taxonomy to remain applicable. This suggests that the benchmark is designed to probe whether a model has learned transferable geometric reasoning rather than sport-specific lexical or visual templates.

4. Evaluation protocol and metrics

The evaluation process for CourtSI-Ext is stated to be consistent with CourtSI-Bench. The same question-answer formats are used; categorical outputs are scored with Accuracy for exact matching; and numerical outputs are evaluated with the same thresholded metric used in CourtSI-Bench when applicable (Yang et al., 10 Mar 2026).

For numerical answer tasks in distance measurement and localization, the paper defines T-MRA (Threshold Mean Relative Accuracy) with threshold set $\mathcal{C} = \{0.5, 0.55, \ldots, 0.95\}$ and $T = 15\text{ cm}$. For localization, however, the supplementary evaluation note states that T-MRA is not used because the outputs are 3D coordinates. Instead, localization is scored with a binary accuracy with a smooth threshold of 30 cm: a prediction is counted correct if the 3D localization error is below this threshold, and otherwise incorrect. The paper notes that this threshold is larger than $15\sqrt{3}$, so it is more lenient than the combined 3D error bound implied by the distance threshold.

The methodological significance of this design is that CourtSI-Ext is not evaluated under a relaxed or ad hoc protocol. By keeping formats, metrics, and annotation conventions fixed, the benchmark isolates cross-sport generalization as the primary variable.

5. Reported empirical results

The main experimental result is a comparison between the base Qwen3-VL-8B model and Qwen3-VL-8B fine-tuned on CourtSI. On CourtSI-Ext, the base model achieves 38.2 overall, while the fine-tuned model achieves 51.4, for an improvement of +13.2 points (Yang et al., 10 Mar 2026).

Subtask	Base Qwen3-VL-8B	Fine-tuned Qwen3-VL-8B
Camera-Object	0.9	70.0
Height	50.0	83.1
Object-Line	31.3	34.6
Object-Object	15.2	62.7
Ball counting	33.3	83.3
Player counting	100	100
Localization	0.0	0.0
Ball-Zone	63.6	63.6
Ball-Player	77.8	66.7
Camera-Player	100	70.0
Player-Zone	21.4	28.6
Player-Player	52.9	44.1
Player-Line	52.4	66.7

The gains are broad but nonuniform. Large improvements appear in Camera-Object, Height, Object-Object, and Ball counting. Some subtasks remain unchanged, such as Player counting and Ball-Zone, while others decline after fine-tuning, including Ball-Player, Camera-Player, and Player-Player. Localization remains 0.0 for both versions in the main table.

The paper also reports other model baselines on CourtSI-Ext. The listed overall scores include GPT-5.2: 55.0, Gemini-3-Pro: 13.5, Gemini-3-Pro parsed: 66.8, Qwen3-VL-235B-A22B: 47.9, and LLaVA-OneVision-72B: 43.3. The parsed Gemini result is especially notable because the paper associates the difference between raw and parsed scores with output-format violations rather than purely latent reasoning failure. The fine-tuned CourtSI model does not exceed the best proprietary parsed result, but it substantially improves over its own base model and thereby supplies the paper’s main evidence for cross-sport transfer.

6. Interpretation, limitations, and diagnostic role

The authors interpret CourtSI-Ext as evidence for partial transfer of court-geometry-grounded spatial reasoning, but not as evidence that the generalization problem is solved. Three points are emphasized. First, positive transfer exists: fine-tuning on CourtSI improves performance on the unseen pickleball benchmark by +13.2 overall. Second, transfer is weaker than the in-domain gain on CourtSI-Bench, where the same adaptation improves accuracy by 23.5 percentage points. Third, localization remains especially difficult (Yang et al., 10 Mar 2026).

The supplementary note sharpens this last point. It states that the adapted model reduces average localization error to 3.9 meters, compared to about 6 meters from other baselines, but that this reduction does not translate into strong accuracy gains under the benchmark’s correctness criterion. This indicates that approximate geometric competence does not necessarily suffice for thresholded spatial QA.

CourtSI-Ext therefore occupies a specific diagnostic niche. Its importance does not derive from scale; the paper explicitly characterizes it as “a small yet valuable benchmark for broader community validation.” Its role is to test whether a VLM has learned transferable metric and relational reasoning anchored in court geometry, rather than sport-specific priors tied to badminton, tennis, or table tennis. The reported gap between CourtSI-Bench and CourtSI-Ext suggests that current VLMs still rely heavily on such priors and remain sensitive to domain shift even when the unseen sport is structurally similar.