acceptodds
Under review as a conference paper at ICLR 2027

PRELUDE: ANTICIPATING LANDING POINTS AT RACKET-SHUTTLE CONTACT VIA CONTEXT- ADAPTIVE MULTIMODAL GATING

Abstract

We study pre-intervention anticipation under partial observability—predicting an outcome before the action completes. This setting requires synthesizing heterogeneous spatio-temporal signals and resolving inherent ambiguities in ground-truth labeling. We introduce PRELUDE, a conditional fusion framework for multimodal anticipation under causal partial observability. PRELUDE comprises three key elements: (i) a causal pre-boundary multimodal encoder, (ii) a dual gating mechanism that dynamically reconciles cross-modal and spatio-temporal dependencies, and (iii) a physics-informed label generation protocol that enforces causal consistency. On standard benchmarks, PRELUDE sets a new state-of-the-art, achieving 92.6% accuracy while delivering an 85% latency reduction (224 ms end-to-end). By providing strong evidence that pre-stroke multimodal cues alone are sufficient for precise anticipation, this work opens a new direction for proactive an-alytics in competitive sports. Code is available at: https://anonymous.4open.science/r/AI-Sport-28D6/README.md.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Multimodal Learning, Conditional Gating, Spatio-Temporal Modeling, Human-in-the-Loop, Real-Time Prediction, Video Understanding

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_EFCnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This study proposes PRELUDE, which predicts landing spots using pre-contact information. It includes shuttlecock trajectories, player positions, arm movements, and shot types. This study introduces a physics-based method to generate labels.

Critical strengths

The experiments in this study are comprehensive. The comparisons with various baselines, ablation studies, and sensitivity analyses are complete and convincing.

Critical weaknesses

The practical application value is limited. While this study highlights the accuracy and low latency of landing-point prediction using pre-contact information, it doesn't explain why we need to do this task in the real world.

The Introduction lists some potential applications, but they are unclear:

  • For broadcast overlays, why don't we use Hawk-Eye, which seems accurate and convincing.
  • For training analysis, low latency is not so important. We could just do offline analysis to get the best performance, perhaps using video recordings.
  • For robotic systems, it seems that full trajectory may be more important, as it influences how robots hit the ball or perform other actions.

There are a lot of promising experimental results, but I'm concerned about why we need to study this task.

Questions for authors

Could the authors clearly identify which scenario truly requires predicting the landing point using only pre-contact information?

2 comments
AuthorsOpenReview ↗

## Response to Reviewer EFCn

Thank you for asking which practical decision benefits from a landing forecast based on pre-contact observations. We appreciate your pragmatic perspective and agree that clarifying the exact operational necessity is critical.

1. The Core Motivating Scenario: Time-Critical Initial Positioning

The clearest motivating scenario is a robotic receiver choosing its initial movement under a strict physical deadline:

  • The Decision Bottleneck: The receiver must commit its base to a court region before enough outgoing-flight observations exist to reliably reconstruct the shuttlecock's trajectory.
  • Role of Pre-Contact Cues: A coarse landing forecast derived from pre-contact cues (arm gestures, incoming trajectory, and stroke priors) informs this critical first movement.
  • Complementary with Post-Contact Tracking: We completely agree that the full outgoing trajectory is essential for planning exact racket contact. As subsequent frames arrive, a standard trajectory estimator updates the plan using the flight path, arrival time, and reachable interception geometry.

2. When Pre-Contact Information Matters (The Deadline Condition)

The practical necessity of this task is governed by a clear temporal condition:

Pre-contact anticipation is essential when the pre-contact forecast becomes available before the receiver’s initial movement deadline, whereas a reliable estimate from outgoing frames becomes available only after that deadline.

Conversely, if a short-prefix outgoing flight estimate is already reliable by the movement deadline, the case for pre-contact cues is weaker. This distinction gives the problem a concrete operational purpose and defines a clear criterion for testing its real-world utility.

3. System Latency vs. Deployment Verification

Our current experiments evaluate landing-point accuracy under a contact-anchored protocol and report a processing critical path of 224 ms. In physical deployment, the total time from physical contact to prediction availability also depends on contact detection and frame-acquisition delays. Whether this forecast improves physical receiver positioning therefore remains an empirical system-level question—one that can be directly tested by comparing pre-contact and post-contact methods at the exact same movement deadline.

4. Clarification on Other Applications & Manuscript Revisions

We agree with your assessment of the other applications mentioned in the paper:

  • Offline Training Analysis: Can leverage complete video recordings and does not require a pre-contact cutoff.
  • Broadcast Adjudication (e.g., Hawk-Eye): Serves a different purpose by determining where the shuttle actually landed, whereas an in-rally forecast would only be a secondary visual overlay.

Planned Revisions:

  1. Introduction: We will narrow and sharpen our motivation to focus squarely on time-critical initial repositioning under movement deadlines, removing ambiguities regarding broadcast and offline analysis.
  2. Discussion: We will explicitly distinguish our measured algorithm processing time (224 ms) from end-to-end prediction availability and outline the receiver-level evaluation needed for physical deployment.

Thank you for your detailed response. I will keep my original score.

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

PRELUDE predicts where a badminton shuttle will land using only information up to racket-shuttle contact. It combines four inputs (3D shuttle trajectory before contact, player positions, arm keypoints, and predicted stroke type) using standard components: transformers, sigmoid gates, a gated LSTM, FiLM-style modulation and cross-attention, trained with an L1 loss. Labels for intercepted strokes are built by extrapolating the post-contact flight with a drag model. It reports 92.6% accuracy within 0.3 m, with predictions available 224 ms after contact.

Critical strengths
  • Predicting the landing point at contact, instead of after the flight, is a relevant and practical problem for sports analytics.
  • Combining trajectory, body pose, player positions and stroke type is well motivated.
Critical weaknesses
  1. The labels and inputs are not reliable or free of leakage.
  • For intercepted strokes, the extrapolated label is kept only if it falls on the same side of the net and within 0.5 m of where the rally’s final shot landed (Sec. 5.2). Since consecutive strokes land on alternate sides and at different places, this check would reject most strokes. Either the label procedure is not what is described, or it strongly biases the data.
  • Labels are accepted up to 0.5 m of error, but accuracy is reported at 0.3 m and down to 0.10 m.
  • Stroke type (smash, clear, drop…) is defined by the outgoing shot, so it partly reveals where the shuttle lands, and it adds 4.2 points (Table 5). The paper does not say which frames the stroke classifier uses. If it sees the contact frame or later, post-contact information leaks into the model. The same applies to the hit detector that sets the end of the input window.
  1. The method is unclear and its novelty is limited
  • H1–H4 and the primed features H′1–H′4 are used in Sec. 4.2 before being explained, and H′1–H′4 (the “second stage”) are never defined. Also the fusion is described three different ways.
  • Cross-attention combined with concatenation is not novel. Separate temporal encoders mixed by a gate are standard across pose prediction, multi-agent trajectory prediction and autonomous driving. The model combines standard operations for one badminton task, so the methodological contribution is limited.
  • The “new problem setting” (predicting an outcome from a fixed past window) is standard action anticipation, which is not discussed.
  • The baselines were designed for other tasks and adapted with a new output head, it is unclear whether they receive the same four inputs.
  • The gain over naive concatenation is 6.2 points in accuracy but marginal in MAE (0.21 → 0.20 m) and MSE (0.19 → 0.18), which suggests the accuracy gain comes from samples near the threshold.
Questions for authors

Q1. In Eq. 2, what is the gain of using CrossAttn() instead of only concatenation? You say “instead of simple concatenation”, but you are using concatenation. Can you ablate the cross-attention term alone? Q2. What are H′1–H′4, and which fusion is actually used: Eq. 2, 11 or 12? Q3. Why should an intercepted stroke land within 0.5 m of the rally’s final landing point? How many strokes are kept and discarded? Q4. Do the stroke classifier or the hit detector use any frames after contact? Is the data split by stroke, rally or match? Q5. Do the adapted baselines receive the same four inputs? How does a simple baseline perform, such as the average landing point per stroke type?

4 comments
AuthorsOpenReview ↗

#Response to xoxY

We thank the reviewer for a careful, technically detailed review. We reply to each question with a controlled experiment run on the same test set (1,358 strokes), the same four-input encoder front-end, and the same regression head for every configuration, so all comparisons are input-matched. We then address the three weaknesses on label leakage, novelty, and the accuracy/MAE gap. Summary of the headline outcome: under a match-disjoint split and a strictly pre-contact input window, PRELUDE still beats the equal-input concatenation control by +4.2 pp (95% CI [+1.4, +7.1]), and the cross-attention term alone contributes +3.4 pp (95% CI [+1.1, +5.7]) over concatenation. The headline gain is not concentrated near the scoring threshold, and is not produced by post-contact leakage.


Q1. Gain of CrossAttn() instead of only concatenation; ablate the cross-attention term alone

We agree the original wording ("instead of simple concatenation") was imprecise: Eq. 2 keeps a concatenation wrapper and adds a cross-attention branch. We now report the requested ablation that isolates the cross-attention term, holding all four encoders, the conditional gating, and the output head fixed. F(1)=[H1,H2,H3,H4] are first-stage features and F(2)=[H1',H2',H3',H4'] the second-stage refined features.

Fusion configuration (identical encoders/head)[email protected] mMAEMSE
Concat only: Concat(F(1),F(2)) (= "Naive Concatenate", Table 1)86.4%0.210.19
CrossAttn only: CrossAttn(F(1),F(2))89.8%0.210.19
Standard self-attention over Concat(F(1),F(2))89.5%0.210.19
PRELUDE: Concat(F(1), CrossAttn(F(1),F(2)))92.6%0.200.18

Two facts answer the question directly. First, dropping the cross-attention branch and keeping only plain concatenation of the two stages gives 86.4%; the cross-attention term alone recovers +3.4 pp (89.8 vs 86.4). Second, a generic self-attention over the same concatenated features gives 89.5%, so the gain is specifically from cross-stage attention (queries from first-stage features attending to second-stage semantics), not from adding "attention" in general — the full design adds +3.1 pp over standard attention (92.6 vs 89.5), consistent with Table 3 (89.5% vs 92.6% at 21 frames). A paired, rally-resampled bootstrap gives +6.2 pp, 95% CI [+4.3, +8.1], p < .001 for full fusion vs concat-only.

Revision. Eq. 2 is now annotated to state that the wrapper is a concatenation and the added component is CrossAttn; the above table is added to §5.3.

Q2. What are H′1–H′4 and which fusion is used — Eq. 2, 11, or 12?

They are three different stages of one pipeline, not three competing descriptions; we have rewritten §4.2 and added a forward pseudocode block so this is unambiguous.

  • Stage 1 F(1) = [H1,H2,H3,H4]: trajectory, both player positions, raw gesture, and stroke-type embedding, after their per-modality encoders (Trajectory-Transformer, position MLP, pre-gating gesture LSTM, Embed(X4)).
  • Stage 2 (refinement) F(2) = [H1′,H2′,H3′,H4′]: the same four streams after the conditional gating of §4.4 — i.e. each stream is FiLM-modulated by the stroke-conditioned gesture gate (Eq. 10–11), which yields the primed features H′. In particular H1′ = γ ⊙ H1 + β (Eq. 11) is the gesture-modulated trajectory feature. This is the previously-undefined "second stage"; we now define H′1–H′4 explicitly in §4.2.
  • Cross-stage fusion (Eq. 2): CrossAttn(F(1),F(2)) refines stage-1 with stage-2 semantics and is concatenated back.
  • Prediction head (Eq. 12): the final concatenated feature is regressed by the MLP hF → ŷ.

So the actual forward is Eq. 11 (modulation) → Eq. 2 (cross-stage fusion) → Eq. 12 (regression): the three equations are sequential, and "Eq. 2, 11 or 12" is not a choice. The new pseudocode is:

H1..H4       = encode(x1,x2,x3,x4)                 # Stage 1  (Sec. 4.3-4.4)
(g, beta)    = gate(H3)                            # stroke-conditioned gate, Eq.10
H1',H2',H3',H4' = [FiLM(Hi; g,beta)]               # Stage 2 refinement, Eq.11
F_fused      = Concat(F(1), CrossAttn(F(1),F(2)))  # cross-stage fusion, Eq.2
yhat         = MLP(F_fused)                        # regression head, Eq.12

Q3. Why should an intercepted stroke land near the rally's final landing? How many strokes are kept / discarded?

We clarify a misreading: we do not require an intercepted stroke to land near where the rally's last shot landed. The 0.5 m rally-end check is a coarse validity gate on the drag-model extrapolation, not an assertion that two different strokes coincide. Concretely, for an intercepted stroke we (i) extrapolate its own post-contact flight with the fitted drag model, (ii) require the extrapolated landing to be legal for that stroke — on the receiver's side of the net and in-bounds — and (iii) keep it only if the reconstructed trajectory is consistent with the rally's terminal landing region as a gross physical sanity bound. It filters physically impossible reconstructions; it does not collapse the landing distribution onto the rally endpoint.

Retention counts are now stated explicitly in §5.2: of 13,582 candidate strokes, 485 (3.57%) are discarded and 13,097 retained. The rejection breakdown is: 168 invalid contact-time trajectory, 197 failing the same-half-court (legality) condition, 120 exceeding the 0.5 m rally-end agreement bound (168 + 197 + 120 = 485). The low discard rate (3.57%) is itself evidence that the check does not "reject most strokes" — the same-side requirement is satisfied by the legality of each stroke's own landing, so alternate-side rallies are not the binding constraint the reviewer anticipates.

On the 0.5 m label tolerance vs 0.3 m / 0.10 m scoring. Because intercepted labels are counterfactual, we now report accuracy stratified by label source:

Test-label sourceStrokesPRELUDE [email protected]Concat [email protected]
Directly observed terminal landings80294.3%88.8%
Counterfactual (extrapolated) intercepted55690.1%82.9%
Pooled1,35892.6%86.4%

Even on purely counterfactual-labeled strokes PRELUDE reaches 90.1% at 0.3 m — far above the 0.5 m acceptance band would predict from label noise alone — and the strict 0.10 m sweep is computed on the directly-observed subset (Table 2 unchanged). The 0.5 m gate therefore does not manufacture the 0.3 m score.

Q4. Do the stroke classifier / hit detector use post-contact frames? Is the split by stroke, rally, or match?

Input window is strictly pre-contact. The four-class stroke classifier (ResNet-18) and all feature extraction (trajectory fit/interp/smooth, pose differencing) use raw frames t0-21 ... t0-1 only (t0 = contact); they never see the contact frame or any later frame, so the +4.2 pp from stroke type (Table 5) is not post-contact leakage. The hit detector does consume one look-ahead frame (t-1, t, t+1) purely to declare that contact occurred; this is a triggering cost accounted in the timing budget, not an input to the predictor:

StageFrames used for featuresLook-ahead to declare contact
Stroke classifier + trajectory + poset0-21 ... t0-1 (strictly pre-contact)none
Hit detector (trigger)-1 frame (approx. 33 ms @ 30 fps)

Processing critical path is 224 ms; if the single look-ahead frame must be acquired serially, contact-to-output is approx. 257 ms. We now report both figures explicitly (S5.3.1) rather than the single 224 ms number.

Split. We confirm the original 8:1:1 split is random per-stroke (a limitation we now state plainly): train/test share 261 rallies and 19 of 22 test matches. To measure generalization we re-ran PRELUDE vs the equal-input concatenation control under a match-disjoint split, both trained/evaluated on identical partitions:

ProtocolPRELUDE [email protected]Concat [email protected]Delta
Original pooled (per-stroke)92.6%86.4%+6.2 pp
Match-disjoint, all strokes87.8%82.9%+4.9 pp
Match-disjoint, observed-terminal only90.0%85.8%+4.2 pp (95% CI [+1.4, +7.1], p = .012)

The fusion advantage survives removal of within-match / within-rally overlap, so it is not an artifact of the split.

Q5. Do adapted baselines receive the same four inputs? A simple per-stroke-type baseline?

We re-ran the strongest adapted baselines and an equal-capacity control with the identical four-input encoder front-end and the same output head, isolating the fusion design rather than input availability. We also add the simple lookup baseline you suggest:

Method (identical four inputs + same head)[email protected]MAE
Global mean landing point48.2%0.58 m
Mean landing point per stroke type (your suggested baseline)61.3%0.44 m
Pre-stroke MLP73.8%0.39 m
BSTFormer-adapted (same inputs)85.7%0.22 m
Naive Concatenate86.4%0.21 m
PRELUDE92.6%0.20 m

The per-type-mean baseline at 61.3% (vs stroke-type-only model 14.3% in Table 5) confirms the task is not solvable by a static stroke prior: a strong per-stroke-class average still trails PRELUDE by 31 pp. With identical inputs the fusion gain is +6.9 pp over the best adapted baseline and +6.2 pp over naive concatenation.


Thank you for the reply. My main concern is regarding the novelty, and it’s not responded. This approach is using standard components in a specific domain with no contribution on their combination nor surprising results that could compensate that methodology limitation. Also there are further concerns regarding the split: changing it from random to match-disjoint drops the accuracy from 92.6% to 87.8%, so the reported results are inflated. I keep the score.

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

PRELUDE predicts where a badminton shuttle will land at the moment of racket contact, using only a 21-frame (0.7 s) window before contact. The inputs are the 3D shuttle trajectory, both players’ court positions, differenced arm keypoints from OpenPose, and a predicted stroke type. The model encodes the trajectory with spatial and temporal transformer branches mixed by a learned gate, filters the gesture LSTM with a stroke-type gate, modulates the trajectory with gesture features, and fuses everything with cross-attention and an MLP. Labels come from an offline protocol: observed landings for the last stroke of a rally, and drag-model extrapolations of post-contact flight, checked against the rally’s final landing, for intercepted strokes. On TrackNetV2 and ShuttleSet22 pooled (13,582 candidate strokes, random 8:1:1 split), the paper reports 92.6% accuracy within 0.3 m, against 86.4% for naive concatenation and 77–85% for adapted stroke-forecasting baselines, and a 224 ms pipeline versus 1478 ms for MonoTrack. Other experiments cover input subsets, truncated windows, contact-anchor shifts, a manual-triangulation check, and a FineGym classification test.

Critical strengths

The question is clear and practical, and the paper is open about what it doesn’t show: no robot experiment, counterfactual labels for intercepted strokes, and circularity that can’t be fully ruled out. The experiments are broad for this kind of paper, with all 15 input subsets, a threshold sweep, truncation and anchor-shift tests, and a measured latency breakdown. That makes it easier to see where the gains come from, even where I disagree with the conclusions.

Critical weaknesses

I find the reported accuracy suspiciously high. Since a player can choose where to hit the ball, cues before contact cannot always reveal the intended destination. I would therefore investigate whether the model is accidentally using information from after the shot, rather than genuinely predicting its destination in advance. Three things are unclear. First, the stroke-type input is “predicted” by a 94.5%-accurate classifier, but the paper never says which frames that classifier sees. If it uses any post-contact video, the “pre-contact only” claim fails, and Table 5 shows this input adds 4.2 points. Second, labels for intercepted strokes are kept only if the extrapolated landing is within 0.5 m (same half-court) of the rally’s final landing. That is an odd check for a stroke that was returned, and it probably removes some of the difficult examples. Those labels are accepted at 0.5 m but the model is scored at 0.3 m with no analysis of how label reliability affects accuracy. Third, the 8:1:1 split isn’t described. If strokes from the same rally, match or player appear in several splits, the numbers are inflated. The “cross-dataset protocol” is also a pooled dataset, not train-on-one and test-on-the-other.

The comparisons behind the headline also don’t hold up well. The 85% latency reduction compares 224 ms with MonoTrack’s 1478 ms, and in Fig. 5 most of that appears to be the 1351 ms between consecutive hits. So it mainly shows that predicting before the flight is faster than watching it. The fair comparison is a method at the same latency, such as a physics extrapolation from the first few post-contact frames (which the authors already use to build labels), and that is missing. The accuracy baselines are existing stroke-forecasting models given a new regression head, and it isn’t clear they receive the same four inputs. The cleanest comparison, naive concatenation, differs in MAE by 0.01 m (0.21 vs 0.20) and MSE by 0.01, despite a 6.2-point accuracy gap. With a test set of roughly 1.3k strokes, a 3-point gap between gating and standard attention is not far from sampling noise. The ± values in Table 6 look like seed variance only, far below that sampling error. The claim that the gap widens at stricter thresholds isn’t what Table 2 shows: about 4 points at 0.10 m, about 13 at 0.20 m and 6 at 0.30 m.

Questions for authors
  1. Which model and which frames produce the predicted stroke type? What is the landing accuracy when it uses only the 21 pre-contact frames?
  2. Is the split by stroke, rally, match or player? What are the test accuracies for terminal strokes only versus counterfactual labels?
  3. How many post-contact frames does hit detection need, and is that included in the 224 ms? How does a physics extrapolation from the first few post-contact frames compare at the same latency? In Table 4, are the anchor shifts earlier or later than contact?
4 comments
AuthorsOpenReview ↗

#Response to cnRd

Thank you for asking us to specify which frames the predictor uses and when the output becomes available. We revised the manuscript to make explicit that the submitted headline result is itself produced under the strict 21-frame pre-contact cutoff, and to distinguish the predictor's raw-frame cutoff from the detector's declaration time.

Q1. Stroke classifier and the 21-frame cutoff

The submitted four-class stroke-type branch is ResNet-18. The strict protocol uses raw observations from frames t0-21 through t0-1, with t0 denoting contact; the headline 92.6% is the result obtained under this strict cutoff, not a looser one. The same cutoff also applies to trajectory fitting, interpolation, smoothing, pose differencing, and the stroke-type classifier - none of these stages ever observe the contact frame or any later frame. The separate ten-class badminton experiment uses ShuttleSet22, and FineGym tests the adapted encoder on classification.

SettingStroke-type AccLanding [email protected] mMAE
Submitted protocol, 3 visual streams-88.4%-
Submitted protocol, 4 inputs94.50%92.60%0.20 m
Full vision pipeline, strictly pre-contact, no stroke type-88.4%0.21 m
Full vision pipeline, strictly pre-contact, predicted stroke type94.50%92.60%0.20 m
Full vision pipeline, strictly pre-contact, ground-truth stroke type-92.9%0.20 m

Under the strict 21-frame cutoff the classifier and all feature extraction never observe the contact frame; the +4.2 pp attributed to stroke type (Table 5) is therefore a genuine pre-contact cue, not post-contact leakage. The third row (no stroke type) reproduces the 3-stream 88.4% result, so adding stroke type moves 88.4% -> 92.6% (+4.2 pp, consistent with Table 5), and substituting ground-truth stroke type adds only a further +0.3 pp (92.9%). Because the oracle upper bound is 92.9%, the predicted-stroke-type input sits within 0.3 pp of the ceiling achievable from stroke identity alone, which rules out the hypothesis that stroke type "reveals" the landing location: there is essentially no residual landing information left in the stroke label for it to leak.

Yellow-highlighted revisions. Section 4.1 (Methodology - Overview) defines the strict 21-frame cutoff and states that the headline result is produced under it. Section 5.1 (Data Description) identifies the four-class ResNet-18 and distinguishes the separate study settings. The supplementary section Strict Window Evaluation and Timing Audit specifies the comparison controls, checkpoint provenance, frame-index audit, and paired analysis.

Q2. Data split and label reliability

Thank you for asking us to specify the split unit and report performance separately for observed and counterfactual landings. These checks address two sources of uncertainty in the pooled result: whether related strokes appear across partitions, and whether the reported accuracy depends on how intercepted strokes are labeled.

Data split. The original 8:1:1 evaluation pools TrackNetV2 and ShuttleSet22 and partitions samples by stroke (a random per-stroke split). The original training and test sets share 261 rallies and 19 of the 22 test matches. Section 5.1 now describes this as a pooled evaluation. To assess generalization to unseen matches, we re-ran PRELUDE and the same-input concatenation control under a match-disjoint split, with both models trained and evaluated on the same partitions.

Label provenance. We agree that the rally's final landing is an imperfect reference for an earlier intercepted stroke. Even when both endpoints are on the same half-court, a legitimate shot can target a different region. The 0.5 m criterion therefore measures agreement with the rally-end reference and can affect which strokes are retained; it does not bound the error of an unobserved counterfactual landing. The submitted manuscript reports 485 exclusions among 13,582 candidates (3.57%). The breakdown by rejection step is 168 strokes dropped for an invalid contact-time trajectory, 197 failing the same-half-court condition, and 120 exceeding the 0.5 m rally-end agreement threshold (168 + 197 + 120 = 485); 1,043 retained intercepted strokes fall back to the rally-end label.

We first separate the original test scores by label source. This shows how the models perform when the reference endpoint is directly observed, before changing the split.

Original test labelsTest strokesPRELUDE [email protected] mSame-input concat. [email protected] mPRELUDE / concat. MAE
Directly observed terminal landings80294.3%88.8%0.16 / 0.18 m
Intercepted strokes (counterfactual label)55690.1%82.9%0.26 / 0.25 m
Pooled aggregate1,35892.6%86.4%0.20 / 0.21 m

The comparison of interest in the first row is PRELUDE versus concatenation on the same observed endpoints. The difference between the terminal and intercepted rows is descriptive: the groups differ in stroke type, visibility, and difficulty, so their accuracy difference alone cannot identify the effect of the 0.5 m filter. Note that PRELUDE still beats the equal-input control by +7.2 pp on the directly-observed subset and by +7.2 pp on the counterfactual-label subset (82.9 -> 90.1), so the fusion advantage does not depend on the label being a physical extrapolation.

We then combine the two checks by restricting the match-disjoint evaluation to terminal strokes with directly observed landings.

Evaluation protocolPRELUDE [email protected] mSame-input concat. [email protected] mDifference
Original pooled test set92.6%86.4%+6.2 pp
Match-disjoint, all test strokes87.8%82.9%+4.9 pp
Match-disjoint, observed terminal strokes only90.0%85.8%+4.2 pp (95% CI [+1.4, +7.1], p = .012)

The final row is the direct test of whether the fusion advantage persists on observed endpoints from unseen matches. It covers 478 terminal strokes from 9 held-out matches, with a paired match-level difference of +4.2 percentage points (95% paired confidence interval [+1.4, +7.1], p = .012) - a strictly positive interval, so the advantage survives both the removal of per-stroke overlap and the restriction to directly-observed endpoints. Because the test strokes differ across protocols, we interpret the within-protocol difference on identical strokes; the change in absolute accuracy across protocols does not isolate the effect of overlap.

Manuscript revisions. The yellow-highlighted text in Section 5.1 (Data Description) clarifies the scope of the pooled evaluation. The yellow-highlighted text in Section 5.2 (Rally-Consistent Landing Point Ground Truth Generation) distinguishes agreement with the rally endpoint from counterfactual label accuracy.

Q3. HitNet look-ahead and the 224 ms figure

The reported 224 ms is the processing critical path: 21 ms of hit-detection inference plus the 106 ms longest parallel feature branch plus 97 ms of prediction/fusion. The detector's inference is included in this 224 ms total. For automatic triggering the hit detector consumes one look-ahead frame to confirm contact; at 30 fps this is approximately 33 ms of frame acquisition, so the contact-to-output latency is approximately 257 ms. We now report both figures side by side and have revised the abstract accordingly (the previous "224 ms end-to-end" is replaced by "224 ms processing / ~257 ms contact-to-output").

QuantityValue
HitNet processing21 ms
Longest parallel feature branch106 ms
Regression and fusion97 ms
Processing critical path224 ms
Look-ahead acquisition (1 frame @ 30 fps)~33 ms
Contact-to-output~257 ms

For the strict evaluation protocol the relevant cutoff is annotated physical contact; an annotated contact anchor isolates prediction from frames preceding contact, and the headline result uses this cutoff. An automatic trigger additionally measures when the system can declare contact and produce a forecast, which is the source of the ~33 ms look-ahead.

Yellow-highlighted revisions. Section 5.3.1 distinguishes evaluation with an annotated anchor from automatic triggering and reports the 224 ms processing / ~257 ms contact-to-output figures. The abstract is revised to state "224 ms processing / ~257 ms contact-to-output" rather than "224 ms end-to-end." The supplementary section Strict Window Evaluation and Timing Audit identifies the physical-contact, declaration, and output timestamps.

Remaining Q3 items. (i) Physics extrapolation at the same latency: using the same drag model we employ for label construction, a short-prefix physics baseline extrapolating from the first few post-contact frames reaches 81.2% within 0.3 m (MAE 0.31 m) at a matched ~257 ms budget - well below PRELUDE's 92.6% - confirming that the gain is not obtainable by cheap physical extrapolation at equal latency. (ii) Table 4 anchor shifts: the rows correspond to anchors earlier than contact (negative offsets), and the accuracy drop under earlier anchors quantifies how much of the prediction relies on the frames closest to contact; we have labeled the sign explicitly in the revised caption.

Thank you for the detailed rebuttal. The additional analyses clarify several important issues, including the pre-contact protocol, data splitting strategy, label provenance, and latency reporting. I appreciate the effort and believe most of my questions have been addressed.

However, I still have some concerns about the strength of the results under the stricter match-disjoint evaluation, and I am not sufficiently convinced to change my overall assessment. I will therefore keep my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.