acceptodds
Under review as a conference paper at ICLR 2027

UniPitch: Parametric Reconstruction of Soccer Matches from Broadcast Videos

Abstract

Reconstructing soccer matches from broadcast videos requires identifying players across changing views and recovering their positions on the pitch. Existing pipelines estimate player tracks and camera geometry separately, then combine their outputs through projection. This separation leaves reconstructed player positions as downstream outputs, limiting their use as supervision for camera estimation. We present UniPitch, a framework for parametric reconstruction of soccer matches that represents persistent player identities, evolving pitch positions, and camera states within a common formulation. A differentiable geometric mapping allows two-dimensional player position errors to propagate back to camera parameters, enabling reconstruction supervision to guide the geometry itself. UniPitch learns persistent identity associations through appearance memory that outlives short-term tracks, models camera evolution using visual and temporal evidence, and refines player positions with identity-linked histories and geometric context. On SoccerNet-GSR, UniPitch achieves 73.30% P-HOTA, improving over the best decoupled baseline by 4.80%, with gains in both pitch-space matching and identity association. Ablation studies support the contributions of persistent identity memory, temporal camera modeling, and differentiable trajectory refinement.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
soccer game state reconstruction, sports video understanding, camera calibration, multi-object tracking

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rnbjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

UniPitch reconstructs soccer matches from broadcast video. For each visible player it predicts an identity slot and a 2D pitch position. It builds on a decoupled baseline (MOTIP for tracking, PnLCalib for calibration) and adds three modules, trained in three stages:

  • PSIM: a persistent appearance memory (4 prototypes per identity slot) added to MOTIP’s ID logits, to recover players who leave and re-enter the view.
  • TCSM: a temporal camera model supervised through player pitch positions via the homography projection. At inference it is only used to fill frames where PnLCalib is unreliable.
  • SFTR: a bounded residual correction of the projected player positions, using geometry, visual, trajectory and calibration-history features.

On SoccerNet-GSR it improves P-HOTA by 4.8 points over the best decoupled baseline, but stays below rule-augmented methods that use jersey and team cues.

Critical strengths
  • Coupling tracking and calibration is a relevant problem for game state reconstruction. The gains over decoupled baselines are consistent across all metrics, mainly in detection and identity association.
  • The ablations are systematic: each module adds to the result (Table 2), and the SFTR input ablation shows which cues matter (Table 4).
  • The related work covers tracking, field registration and game state reconstruction well.
Critical weaknesses
  1. The main claim (unified, end-to-end, differentiable geometric mapping) is not supported.
  • The “differentiable geometric mapping” is the standard camera → homography projection, used as a loss in TCSM. Reprojection losses on camera parameters are well known (TVCalib)
  • Training is staged, not end-to-end. MOTIP is trained in Stage I and then frozen. PnLCalib uses released weights and is never trained. The gradient from player positions never reaches the calibration that actually places the players.
  • At inference, players are projected with PnLCalib. TCSM only fills unreliable frames, using hand-set rules. Its position supervision adds only +0.56 P-HOTA (Table 3). Most gains come from SFTR and PSIM, which are independent refinements on top of the decoupled pipeline.
  • Only visible players are reconstructed, so “reconstruction of soccer matches” is overstated. Off-screen players, studied in the trajectory-imputation literature [A,B,C], are not addressed.
  • What remains is a set of add-on modules on a decoupled pipeline, with limited novelty (e.g., the memory is close to DeepSORT and MeMOTR).
  • SFTR operation is not properly explained. Which are the four branch features?
  1. The experimental evaluation is not convincing, and key details are missing.
  • UniPitch is below existing methods (73.30 vs 78.96 P-HOTA for KISTGSR). Not using jersey or team cues is a design choice, and adding them is never tested.
  • It uses its own metric (P-HOTA), and the other methods’ numbers were recomputed without explaining how.
  • The Table 2 baseline (66.86) doesn’t match MOTIP + PnLCalib in Table 1 (68.50).
  • Simple controls are missing: MOTIP with a longer history instead of PSIM, direct camera supervision using the available ground truth, and an ablation on the number of prototypes (K = 4).
  • The method is hard to follow. Fig. 1 blocks are not defined in the text, and the paper does not explain what q_i is, how the memory is initialized and updated, or how temporal processing works (autoregressive or sliding window). The FPN is not cited.

[A] Omidshafiei, Shayegan, et al. "Multiagent off-screen behavior prediction in football." Scientific reports 12.1 (2022): 8638. [B] Xu, Yi, and Yun Fu. "Sports-traj: A unified trajectory generation model for multi-agent movement in sports." International Conference on Learning Representations. 2025. [C] Capellera, Guillem, et al. "Jointdiff: Bridging continuous and discrete in multi-agent trajectory generation." International Conference on Learning Representations. 2026.

Questions for authors
  • Q1. SoccerNet-GSR provides ground-truth camera annotations. Why supervise TCSM only through player positions, and how does this compare with direct camera supervision?
  • Q2. In what fraction of test frames is PnLCalib flagged as unreliable, so that TCSM is actually used? Are TCSM’s camera estimates more accurate than PnLCalib’s on those frames?
  • Q3. How were the rule-augmented results in Table 1 (taken from other papers) obtained under P-HOTA, and on which split?
  • Q4. Why does the Table 2 baseline (66.86) differ from MOTIP + PnLCalib in Table 1 (68.50)?
  • Q5. What is q_i? How is the memory initialized and updated, and how does performance change with K? Does PSIM beat MOTIP with a longer history window?
  • Q6. Is inference online or based on sliding windows? Can you report runtime and results over several seeds?
3 comments
AuthorsOpenReview ↗

Response to rnbj - Part 1

Thank you for your careful review.

W1 · Parametric trajectory reconstruction

We would like to clarify that UniPitch is not a set of independent add-on modules.

  1. Unified parametric modeling is the contribution. Persistent identities link pitch positions over time, and camera states connect them to image observations. This geometry lets pitch-position labels supervise temporal camera state modeling (TCSM), whose motion guides correction of unreliable projections.

  2. Staged training retains geometric supervision. We agree that training is staged and projection is standard, and have corrected the end-to-end description. However, position errors update TCSM through projection with the tracker and PnLCalib fixed.

  3. TCSM improves reconstruction through camera correction and position supervision. Table 2 shows that adding TCSM to PSIM+SFTR improves P-HOTA from 71.61 to 73.30, demonstrating the benefit of TCSM-guided correction. Furthermore, Table 3 shows that enabling pitch-position supervision for TCSM improves P-HOTA from 72.74 to 73.30.

Q1 and Q2 · Camera supervision and calibration

  1. Q1: Position supervision directly optimizes reconstruction. SoccerNet-GSR derives player pitch positions by projecting annotated bounding-box bottom centers using estimated camera geometry. However, the publicly available Labels-GameState.json file, available for downloading from the official website, provides player pitch positions and pitch-line annotations but does not include explicit camera parameters or homography matrices. Therefore, direct camera supervision requires additional target construction.

    For comparison, we fitted camera targets from box-bottom centers and pitch labels via RANSAC, homography decomposition, and nonlinear refinement. Pinhole fitting assumes a centered principal point, equal focal lengths, and negligible distortion. With other settings fixed, we averaged Smooth L1 loss over seven fitted parameters. The results are reported in the table below.

    TCSM supervisionP-HOTAP-DetAP-AssAP-LocAP-IDF1
    Fitted camera targets72.9582.3264.6893.3979.09
    Player pitch positions73.3082.6765.0493.3679.32
  2. Q2: Correction affects 21.12% of frames (7,761/36,750). To evaluate the effectiveness of this correction, we follow the camera calibration evaluation protocol of Magera et al. and report results on these unreliable frames. UniPitch correction reduces mean reprojection error by 45.5% (31.97 to 17.42 pixels) and raises JaC@10 from 44.50 to 50.24.

    MethodJaC@5 ↑JaC@10 ↑CR (%) ↑Mean error (px) ↓Median error (px) ↓
    PnLCalib33.6044.5097.9931.972.39
    UniPitch correction35.5250.24100.0017.422.41
AuthorsOpenReview ↗

Response to rnbj - Part 2

W1 and Q5 · Appearance memory and SFTR

We agree that appearance memory is established.

  1. PSIM links identities and trajectories. DeepSORT uses a descriptor gallery, while MeMOTR updates track embeddings via memory attention. Persistent slot identity memory (PSIM) adjusts MOTIP's short-term identity logits via slot-indexed prototypes. These identities organize SFTR histories.

  2. Q5: role of qi\mathbf{q}_i and memory initialization and updating. As detailed in the revised paper, qi\mathbf{q}_i is the Deformable DETR feature, which is served for ID prediction, prototype matching and memory updating. A slot's first KK accepted features initialize its prototypes. Later features update the closest prototype by cosine similarity via EMA.

  3. Q5: Ablations support persistent memory. With a fixed 60-frame window, PSIM raises P-AssA from 58.04 to 61.57 (Table 2). For KK ablation studies, K=4K=4 gives the highest P-HOTA, P-AssA, and P-IDF1 as illustrated in the table below. We have not tested memory-matched longer-history MOTIP.

    KKP-HOTAP-DetAP-AssAP-LocAP-IDF1
    172.9282.7364.3193.4079.03
    273.0682.9064.4393.4379.14
    373.0682.6464.6393.4179.24
    473.3082.6765.0493.3679.32
    572.7081.8164.6593.2279.08
  4. SFTR/Figure 1: Four branches guide refinement. We agree clarification was needed, so we expanded the caption and added the FPN citation. Four branches encode box/projection geometry, pooled DETR features, same-slot position history, and calibration history. Their concatenation predicts a pitch position residual.

  5. Visibility. Yes, we reconstruct visible players, as clarified in the abstract and limitations.

W2, Q3 and Q4 · Evaluation

  1. Q3: All methods share the same protocol. We reproduce the compared methods with their public Github implementations and evaluate them on the official SoccerNet-GSR test split using the same annotations and our P-HOTA evaluator. We agree that KISTGSR scores higher (78.96 vs. 73.30), while UniPitch gains 4.80 points over the strongest decoupled baseline. Explanation have beend added in the paper.

  2. Q4: Baseline configurations differ. Both share same PnLCalib with different configurations in tracker. The tracker in Table 1(MOTIP + PnLCalib, 68.50) uses default SportMOT config, with 50 identity classes plus a newborn class, whereas Table 2(66.86) uses 25 fixed slots and no newborn class with all proposed modules removed . We have clarified this in the revised paper.

Q6 · Inference, runtime and seeds

Per-frame prediction is online with a causal sliding window, whereas correction is offline because it uses past and future anchors. We have clarified the inference details in Appendix C.

The full pipeline averages 230.88 seconds per 750-frame clip (3.25 FPS) over 49 clips on one A100. We report point estimates because multi-seed evaluation has not been run.

I want to thank the authors for the effort in the rebuttal, for revising the end-to-end claims, and for resolving my doubts about the camera annotations, the baseline configuration and the evaluation protocol. However, I still see this approach as a well-engineered pipeline for GSR with limited methodological novelty. The most original idea, supervising the camera model with player positions, is only used to correct frames where PnLCalib fails, and the approach remains below the rule-augmented methods. I keep my score.

Reviewer_ptLqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper creates a framework for parametric reconstruction of soccer games from broadcast videos. They augment short-term trackers with a persistent memory to facilitate long-term reID, and perform trajectory refinement and camera calibration using pitch position localization losses.

Critical strengths
  1. The problem is well-motivated and fairly clear.
  2. The method meticulously integrates various modules together to create a football game reconstruction algorithm without using jersey colours, jersey numbers and other team affiliation information.
  3. The comparisons to SOTA methods are rigorous.
Critical weaknesses
  1. The methodology is not explained clearly. I do not understand the MOTIP logit part in PSIM at all. What does that even do, and why is that logit computed, if it is not being used for memory updation or to find out the slot?
  2. Since PSIM augments MOTIP's short-term identity association with persistent appearance memory, a memory-budget-matched comparison against MOTIP with an extended temporal context and longer track-retention horizon is necessary. This would help establish whether the improvements in long-term identity association stem from the proposed memory mechanism or simply from retaining a longer history of track features.
  3. Also, the differentiable end-to-end aspect of the model is completely nullified by the proposed training strategy. Training just the detection systems initially, then freezing detection and training PSIM, and then training SFTR and TCSM while freezing the other modules converts your pipeline to a multi-stage pipeline that is not end-to-end. The ablations section states that the performance improvement upon using PSIM and SFTR improves P-HOTA which demonstrates the complementary benefits, but that is still not entirely convincing as PSIM could simply provide better detection and tracking results which is directly required for the SFTR component. An ablation focusing just on PSIM and TCSM coud maybe give more information.
  4. Also, in Table 1, it is mentioned that MOTIP+PnLCalib is 68.5 P-HOTA. But in the ablation Table 2, removing all the components only gives 66.86 P-HOTA, even though both of them should be the same technically. Because the pipeline uses MOTIP for short-term identity prediction (L 207) and it uses PnLCalib to estimate homography (L 304), removing all three modules should lead to the same MOTIP+PnLCalib method that should be the same as the number reported in Table 1. If it is the same but still has different numbers, is it sensitive to different seeds? Then a multi-seed run with mean and variance would be better for all the methods as 2% P-HOTA is a significant deviation.
Questions for authors

See Weaknesses

3 comments
AuthorsOpenReview ↗

Response to ptLq - Part 1

Thank you for your detailed comments on training and evaluation.

W3 · Training and the PSIM + TCSM ablation

We agree that training is staged and have corrected the end-to-end training description accordingly. However, we respectfully disagree that staged training diminishes the contribution of our differentiable geometric supervision.

  1. Player positions directly supervise the temporal camera model. Temporal camera state modeling (TCSM) projects annotated image footpoints onto the pitch. Position errors then backpropagate through this projection, training the camera model for trajectory reconstruction. With shared-feature trajectory refinement (SFTR) supervised, adding the TCSM position loss raises P-HOTA from 72.74 to 73.30 (Table 3), supporting the benefit of this supervision.

    This gradient path is retained in the staged training. Stage I trains the detector and short-term identity predictor, and Stage II trains persistent slot identity memory (PSIM) with both frozen. Finally, Stage III freezes the tracker and trains TCSM and SFTR with pitch-position supervision. PnLCalib retains its fixed released weights throughout.

  2. The correction variant improves P-HOTA without SFTR. Since SFTR provides the reliability criterion in the original correction pipeline as stated in Appendix C of the revised PDF, we replace it in this ablation with a TCSM-based criterion. Specifically, we flag frames when Dt>1.5D_t > 1.5 m, where

    \left( \mathbf{p}^{\mathrm{PnL}}_{t, i}-\mathbf{p}^{\mathrm{PnL}}_{t-1,i} \right) - \left( \mathbf{p}^{\mathrm{TCSM}}_{t,i}-\mathbf{p}^{\mathrm{TCSM}}_{t-1,i} \right) \right\\|_2.$$ Here $\mathbf{p}^{\mathrm{PnL}}$ and $\mathbf{p}^{\mathrm{TCSM}}$ are player positions projected by their respective homographies, and $\mathcal{I}_t$ contains players observed in both frames. The variant fuses corrected positions with coarse PnLCalib positions. Adding the correction variant to PSIM raises P-HOTA by 0.25 points, from 68.99 to 69.24. | Configuration | P-HOTA | P-DetA | P-AssA | P-LocA | P-IDF1 | | ------------------ | -----: | -----: | -----: | -----: | -----: | | PSIM | 68.99 | 77.34 | 61.57 | 92.96 | 76.05 | | PSIM + TCSM | 69.24 | 77.58 | 61.83 | 92.73 | 76.28 | | PSIM + SFTR | 71.61 | 80.51 | 63.73 | 92.99 | 78.33 | | PSIM + SFTR + TCSM | 73.30 | 82.67 | 65.04 | 93.36 | 79.32 |
  3. The SFTR residual rule identifies more large-error frames. To investigate this limited improvement, we further compare the reliability of SFTR-based and TCSM-based frame selection. On frames whose median PnLCalib player position error against ground truth exceeds 5 m, recall is 93.16% with SFTR residuals and 59.33% with the displacement rule. We can also observe that adding SFTR with TCSM present raises P-HOTA from 69.24 to 73.30, indicating that the effectiveness of TCSM correction depends on accurately identifying unreliable frames.

W2 · Memory-matched longer-history MOTIP

We agree that the source of PSIM's gain should be examined.

  1. The gain holds at a fixed short-term window. With the window held at 60 frames and PnLCalib fixed, adding PSIM raises P-HOTA from 66.86 to 68.99 and P-AssA from 58.04 to 61.57 (Table 2). This establishes the benefit of adding persistent memory under the stated settings.

  2. Prototypes preserve identity information beyond that window. PSIM does not build on enlarging the temporal window. Each slot retains up to four appearance prototypes across the entire video sequences, supporting identity recovery after earlier observations leave the window. We have clarified this in the revised Section 3.2.

W1 · How the MOTIP logits are used

Yes, the MOTIP logits contribute directly to slot assignment and they are not used for memory updating. We have added relative explanation in the paper.

  1. The logits help to determine the assigned identity. The MOTIP decoder outputs the MOTIP base logit by comparing detection ii's Deformable DETR feature qi\mathbf{q}_i with recent tracklets to score each slot. Then, Equation 3 adds a gated appearance affinity from persistent memory to each MOTIP base logit. Finally, softmax and Hungarian matching use the fused logits to assign a slot.

  2. The detection feature updates the assigned slot's memory. After Hungarian matching assigns detection ii to slot st,is_{t,i}, its feature qi\mathbf{q}_i is used to initialize the slot's appearance prototypes or update the most similar prototype via EMA.

AuthorsOpenReview ↗

Response to ptLq - Part 2

W4 · Table 1 versus Table 2

The scores use different tracker configurations under the same PnLCalib and evaluation protocol. We agree that this distinction needed clarification and have added it to the revised paper.

  1. Table 1 uses MOTIP's SportsMOT configuration. Its 68.50 comes from MOTIP fine-tuned on SoccerNet-GSR with the SportsMOT hyperparameters, 50 identity classes, and a newborn class.

  2. Table 2 holds UniPitch's slot design fixed for ablation. Its 66.86 uses 25 fixed slots and no newborn class, with all three proposed modules removed. Newly observed players enter those slots directly, providing a common configuration for the module ablations. We have not run a multi-seed evaluation.

Thank you for your constructive review. The authors have posted their rebuttal. Could you please review their response to see if you have any remaining questions?

Reviewer_cHzkOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a method for solving a new task called "pitch reconstruction." To do so, the authors propose a framework called UniPitch, which integrates three specific modules (TCSM, PSIM, and SFTR).

The method works by performing a frame-by-frame homographic projection, combining it with a mechanism for (re-)identifying players, and applying a correction to the camera parameters that ensures temporal consistency.

The method is evaluated using a metric called P-HOTA, which ignores semantic information. The results, combined with an ablation study, demonstrate the method’s effectiveness.

Critical strengths

[S1] The calibration part is interesting. It combines a traditional bottom-up approach (detection using the PnLCalib network followed by homography using DLT) with an end-to-end approach (image regression to interpretable camera parameters (pan, tilt, roll, focal length, and position) including temporal filtering; this is a new approach.

[S2] It is also a good idea to allow switching from PnLCalib’s homography mode to a tracking mode, where “relative” homographies between images are calculated using regression rather than an absolute homography from the ground to the image. Supervision is based solely on the players’ positions, which is also interesting; there is no explicit ground truth for the parameters.

Critical weaknesses

[W1] The main originality of this paper lies in the definition of a camera calibration technique. As the authors note (referring to “temporally consistent pitch-space reconstruction”), the approach differs from that used by Broadtrack [Mag+25] and addresses concerns regarding temporal consistency as pointed out by the authors of PnLCalib (see [GA26]). The problem is that this essential component is not evaluated, even though a standard protocol exists (see [Mag+24]).

[W2] Still regarding calibration, it is also unfortunate that there are no results showing parameter changes over time. Furthermore, information is missing to reproduce the network that regresses the camera parameters. Similarly, the following sentence is unclear: “inter-frame camera changes estimated by TCSM are used to propagate neighboring reliable PnLCalib homographies across the unreliable interval and reconstruct corrected homographies.” In fact, the authors could calculate “relative” homographies by assuming that the camera remains fixed or not; this is not specified. Consequently, due to a lack of information, the results are likely not reproducible. Since there is no “Reproducibility” section (which is recommended but not mandatory), this remains a hypothesis.

[W3] I have concerns with the definition of the Pitch-space HOTA (P-HOTA) metric, which deviates from the GS-HOTA metric Giancola et al. [Gia+25] and effectively introduces a new “pitch reconstruction” task that differs from the Game State Reconstruction (GSR) task. This new task can be viewed as an intermediate task between calibration and GSR. The definition of a new task makes any comparison with the “SoccerNet GSR” state of the art impossible.

[Mag+25] Floriane Magera, Thomas Hoyoux, Olivier Barnich, and Marc Van Droogenbroeck. “BroadTrack: Broadcast Camera Tracking for Soccer.” In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). Tucson, Arizona, USA: IEEE, Feb. 2025, pp. 6177–6187.

[GA26] Marc Gutiérrez-Pérez and Antonio Agudo. “PnLCalib: Sports field registration via points and lines optimization.” In: Computer Vision and Image Understanding 267 (Apr. 2026), p. 104712.

[Mag+24] Floriane Magera, Thomas Hoyoux, Olivier Barnich, and Marc Van Droogenbroeck. “A Universal Protocol to Benchmark Camera Calibration for Sports.” In: IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). Seattle, Washington, USA: IEEE, June 2024, pp. 3335–3346.

[Gia+25] Silvio Giancola, Anthony Cioppa, Marc Gutiérrez-Pérez, Jan Held, Carlos Hinojosa, Victor Joos, Arnaud Leduc, and et al. “SoccerNet 2025 Challenges Results.” In: arXiv abs/2508.19182 (2025).

Questions for authors

[Q1] In light of the arguments presented in the paper, I believe it is necessary to evaluate the quality of the calibration step. Given the lack of details and the impossibility to reproduce the results, this would help us understand how UniPitch contributes to solving the GSR task.

[Q2] Since soccer video analysts are interested in the GSR task rather than pitch reconstruction, and given that this is an application research paper, I would like to understand why the authors defined the pitch reconstruction task.

4 comments
AuthorsOpenReview ↗

Response to cHzk - Part 1

Thank you for your constructive comments on calibration and evaluation.

W1 and Q1 · Calibration accuracy

We agree that calibration accuracy should be evaluated directly and have added this evaluation.

  1. The correction improves calibration on the frames where it is applied. On the 7,761 flagged frames (21.12% of the 36,750 test frames), combining PnLCalib anchors with relative motion from temporal camera state modeling (TCSM) reduces mean reprojection error from 31.97 to 17.42 pixels and raises JaC@10 from 44.50 to 50.24. Median error remains similar. Frames outside this subset retain their original PnLCalib homographies.

  2. The comparison measures both field alignment and coverage. JaC measures how well reprojected field markings align with annotations at a 5- or 10-pixel threshold, while calibration completeness (CR) measures the proportion of frames with a valid estimate. Magera et al., CVPR Workshops 2024

    MethodJaC@5 ↑JaC@10 ↑CR (%) ↑Mean error (px) ↓Median error (px) ↓
    PnLCalib33.6044.5097.9931.972.39
    UniPitch correction35.5250.24100.0017.422.41

W2 · Camera motion and reproducibility

We agree that the correction procedure needed a clearer explanation. Details about inference and correction have been added in Appendix C. We also added a Reproducibility Statement linking the method and experimental details.

  1. The camera model combines visual estimates and predicted motion. A feature pyramid network and spatial attention pooling encode each frame. A visual head estimates pan, tilt, roll, focal length, and the 3D camera center from these features. A GRU over a causal 16-frame window predicts the rate of change. A learned gate combines the visual estimate with the previous state updated by the predicted motion.

  2. Reliable frames anchor correction across each flagged interval. During offline correction (Appendix C), we flag a frame if PnLCalib fails to produce a valid homography or if the median distance between coarse and refined player positions reaches 1 m.

    For an interval bracketed by reliable frames, TCSM's frame-specific homographies give relative motion from the left anchor. We use that motion to propagate the pitch projections of a fixed image grid, blend them with interpolation between both anchors, and fit a corrected homography to the resulting correspondences. The calibration evaluation measures the accuracy of the final corrected homographies, thereby assessing the effectiveness of TCSM-guided correction.

AuthorsOpenReview ↗

Response to cHzk - Part 2

W3 and Q2 · Scope and relationship to SoccerNet-GSR

We respectfully disagree that the chosen scope prevents comparisons with GSR methods. At the same time, we agree that full GS-HOTA leaderboard scores address a broader task.

  1. Table 1 compares all methods on the same GSR subtask. We reran the public implementations of the GSR baselines and scored all methods under the same P-HOTA protocol on the same 49 test sequences. P-HOTA uses SoccerNet-GSR's pitch-distance similarity (0.05 at 5 m), retains track IDs for association, and omits role, team, and jersey-number matching. This follows the pitch-only setting explored by Somers et al., SoccerNet-GSR, Sections 5 and 7.3. The revised evaluation section makes this scope explicit.
  2. Pitch-space reconstruction provides an interpretable representation for game analysis. Visible-player positions and consistent identities allow analysts to follow player movements across changing camera views. To support this need, we reformulate this aspect of GSR as a parametric reconstruction problem, explicitly modeling persistent player identities, pitch positions, and camera states within a unified framework. Unlike conventional pipelines that treat pitch positions as downstream outputs of separate tracking and calibration, UniPitch uses annotated pitch positions to supervise camera-state learning through differentiable projection. Its output, identity-consistent player trajectories in pitch coordinates, could directly supports player movement analysis.
  3. Future directions for more independent supervision. SoccerNet-GSR derives pitch-position labels by projecting annotated bbox bottom centers with estimated camera geometry (Section 4.1: Athlete Localization on the Pitch). These labels may therefore inherit errors from player localization and camera calibration. In the future, controlled simulations with known player trajectories and camera states, or synchronized independent measurements, could provide a more independent supervision and further support joint learning of tracking and camera estimation.

Thank you for your constructive review. The authors have posted their rebuttal. Could you please review their response to see if you have any remaining questions?

Reviewer_cHzkOpenReview ↗

Replying to responses

I thank the authors for clarifying my concerns about calibration by providing an evaluation on a selection of images. Since PnLCalib operates on a frame-by-frame basis and the authors use an interpolation mechanism, a more detailed study could shed further light on this essential component—namely, calibration—for sporting events.

Regarding answers to question / Part 2, I understand that the authors are presenting a new task and that there is value in grouping a series of subtasks, such as calibration and player tracking, but I do not see how integrating the components of the GSR task under an overarching framework constitutes sufficient originality; in this regard, I share the opinion of reviewer rnbj.

Despite the clarifications regarding the calibration section, I will maintain my score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall response to the Area Chair and reviewers

We thank the Area Chair and reviewers for their careful reading and constructive feedback.

We would like to further clarify our central contribution: a unified parametric formulation of player trajectory reconstruction that links persistent identities, pitch positions, and temporal camera states. Specifically, we model the physical scene underlying the video, with player positions in pitch coordinates and camera states determining their image projections. Accordingly, our staged implementation uses annotated pitch positions to supervise the temporal camera model through differentiable projection. Simulation with known player and camera states, or synchronized independent trajectory measurements, could extend this formulation toward jointly learning tracking and camera geometry.

In response to the reviewers' comments, we have added targeted experiments and clarified the method, evaluation, and task scope, as summarized below.

  1. Added calibration evaluation (cHzk W1, Q1 / rnbj Q2). On 7,761 frames selected for correction, the final corrected homographies reduce mean reprojection error by 45.5% (31.97 to 17.42 pixels) and improve JaC@10 from 44.50 to 50.24.

  2. Added supervision and ablation experiments (ptLq W3 / rnbj Q1, Q5). We compared player-position supervision with fitted camera targets (73.30 vs. 72.95 P-HOTA), tested the model without SFTR, and varied the number of appearance prototypes. The individual replies provide the complete tables and analysis.

  3. Expanded the method explanation (cHzk W2 / ptLq W1, W3 / rnbj W1, Q5, Q6). We clarified the parametric representation, the training stages and gradient path, and the memory updates. Appendix C now details calibration correction, and we also report full-pipeline runtime.

  4. Clarified task scope and comparisons (cHzk W3, Q2 / ptLq W4 / rnbj W1, Q3, Q4). The abstract and main text now clarify the scope of player trajectory reconstruction. We clarified the common pitch-space evaluation and the tracker configurations behind the two baseline scores. UniPitch achieves 73.30 P-HOTA, 4.80 points above the strongest decoupled baseline.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.