MS-RFD: Multi-Signal Release Frame Detection in Hammer Throw from Reconstructed 3D Trajectories
Ahmed Endris Hasen, Nikolaos Passalis, Tomi Vanttinen, Jenni Raitoharju
cs.CV
Quick verdict
MS-RFD fuses four kinematic cues for automatic hammer-release detection, cutting distance MAE 7.46→4.37 m on 17 throws; promising but small-sample.
Summary
The paper presents MS-RFD, a fully automatic method for detecting the release frame in hammer throw from 3D hammer trajectories reconstructed with the authors' prior dual-view video pipeline (DLT calibration, hammer-head detection, stereo triangulation). Four physics-informed signals — speed dynamics, angular velocity transition, radial expansion from a RANSAC-estimated rotation center, and post-release trajectory linearity (R²) — are normalized within a search window placed around the final rotation cycle, fused with fixed weights (0.40/0.20/0.30/0.10), filtered by biomechanical release-parameter bounds, and finalized via three candidate-selection strategies (weighted mean, representative frame, multi-criteria). On 17 throws recorded by KIHU (2017–2024) at 240/180 fps, evaluation is indirect: release parameters at the detected frame feed a ballistic model whose predicted distance is compared to the measured distance. The best configuration (C4, weighted-mean selection) achieves MAE 4.37 m (MAPE 6.19%) versus 7.46 m for the authors' previous automatic baseline (−41.4%) and 2.95 m for manual selection; ablations show radial expansion gives the largest gain after speed dynamics. Main limitations: evaluation lacks frame-level ground truth, conflating detector, reconstruction, and ballistic-model errors; the dataset is small and private; fixed hand-set weights risk overfitting; no code or significance testing is reported.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A weighted fusion of four kinematic signals (speed dynamics, angular transition, radial expansion, post-release linearity) can automatically locate the hammer release frame from reconstructed 3D trajectories without manual inspection.
- The full C4 configuration with weighted-mean selection reduces throwing-distance MAE from 7.46 m (previous automatic baseline) to 4.37 m (−41.4%, MAPE 6.19%) on 17 throws.
- Ablation results show speed dynamics and radial expansion are the dominant cues, while angular transition and post-release linearity add smaller refinements.
- MS-RFD substantially closes but does not close the gap to manual release-frame selection (2.95 m MAE).
Strengths
Physics-informed, interpretable, training-free design
Each signal (Sec. II-B) is grounded in the rotational-to-ballistic transition — speed rise, angular-change drop, sustained radial expansion from a RANSAC-estimated rotation center, and post-release R² linearity — so the detector needs no annotated training data, an advantage for a niche sport with no public datasets.
Systematic ablation with clear signal attribution
Table I crosses four signal configurations (C1–C4) with three selection strategies (S1–S3); the C2→C3 step (adding radial expansion) yields the largest improvement (best MAE 4.72→4.39 m), directly supporting the paper's claim about which cues matter.
Transparent per-throw reporting including failures
Honest positioning against manual upper bound
Weaknesses
Very small, private evaluation set
Indirect evaluation conflates multiple error sources
No statistical rigor behind the rankings
Hand-set, under-specified hyperparameters with overfitting risk
Baseline set limited to the authors' own prior work
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How accurate is automatic release frame detection in hammer throw?
In this paper, the best automatic method (MS-RFD) yields a predicted throwing-distance mean absolute error of 4.37 m (about 6.2% of a ~70 m throw) on 17 recorded throws, compared with 2.95 m when experts pick the release frame manually and 7.46 m for the authors' earlier automatic baseline. The paper does not report direct frame-level accuracy because no annotated release frames were used.
What signals are used to detect the hammer release moment?
Four kinematic cues computed from a reconstructed 3D hammer trajectory: (1) increasing hammer speed near release, (2) a drop in the rate of velocity-direction change after release, (3) sustained outward radial expansion away from the estimated rotation center, and (4) linearity (R²) of the trajectory segment following the candidate frame. They are normalized, fused with weights 0.40/0.20/0.30/0.10, and filtered by biomechanical plausibility bounds on speed, angle, and height.
Does the paper provide code or a public dataset?
No. The evaluation uses a private dataset of 17 hammer throws collected by the Finnish Institute of High Performance Sport (KIHU) between 2017 and 2024 with dual cameras at 240/180 fps, and the paper mentions no code or data release, which limits reproducibility.
Why does the release frame matter for hammer throw distance estimation?
Release speed, release angle, and release height read at the release instant determine the subsequent ballistic flight and hence the predicted throw distance. If the frame is wrong, these parameters are biased and the predicted distance is off; manually finding the frame is slow and subjective, which motivates automatic detection.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–4 are present only as captions in the extracted text, so the framework diagram, error-bar magnitudes, boxplots, and the qualitative 3D trajectory could not be assessed.
- Some equations are garbled by extraction (absolute-value bars missing in the MAE/MedAE/MAPE formulas); standard definitions were assumed.
- Numeric hyperparameters (window lengths, L, biomechanical bounds) are absent from the text, so reproducibility judgments are based on their omission.
- The evaluation dataset is private, so no result could be cross-checked against public data.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.