REVIEW 4 major objections 3 minor 1 cited by
TOTNet: Occlusion-Aware Temporal Tracking for Robust Ball Detection in Sports Videos
T0 review · 4 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash
Pith's one-line read TOTNet tracks balls through full occlusion, cutting positional RMSE from 37.30 to 7.19 across four sports datasets.
desk verdict Real contributions in the dataset and architecture, but the headline accuracy gains are uninterpretable without the full evaluation protocol. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is a 3D convolutional network operating over a short temporal window of frames, trained with a visibility-weighted loss and occlusion augmentation. The visibility-weighted loss up-weights frames with known ball positions so uncertain or missing labels do not dominate the gradient; occlusion augmentation randomly hides the ball during training so the model must predict trajectories through invisible spans. The 3D convolutions are what let the network exploit motion and trajectory priors across the window.
What would settle it
Run the earlier trackers on the same train/test splits and the same occlusion label definitions as TOTNet, with identical post-processing; if their RMSE on fully occluded frames approaches 7.19 or the accuracy gap toward 0.80 closes, the claimed superiority is protocol-driven rather than algorithmic.
Extended reading notes
Core claim
The paper's central claim is that occlusion-resilient ball tracking is best achieved by giving the model explicit temporal context rather than a stronger per-frame detector. TOTNet treats a short window of frames as a 3D volume, weights the loss by whether the ball is visible, and trains on clips where occlusion is artificially inserted. The reported result is a large drop in tracking RMSE—from 37.30 to 7.19—and a jump in fully occluded frame accuracy from 0.63 to 0.80 on four datasets in tennis, badminton, and table tennis. The paper also introduces TTA, a new occlusion-rich table tennis dataset with 9,159 samples, 1,996 of them occlusion cases, drawn from professional-level matches.
Load-bearing premise
The comparisons assume earlier methods were evaluated on the same train/test splits, the same occlusion definitions, and the same post-processing, so the reported RMSE drop and accuracy gain reflect the algorithm rather than evaluation-protocol differences.
Editorial extensions
If this is right
- Offline ball tracking in tennis, badminton, and table tennis can preserve ball position through full occlusion, supporting more reliable automatic line calls and event detection.
- The visibility-weighted loss offers a way to train trackers when ground-truth ball positions are available for only some frames in a clip.
- Occlusion augmentation can be applied to existing tracking datasets without new data collection, making the approach transferable to other sports with labeled footage.
- The new table tennis dataset with 9,159 samples, including 1,996 occlusion cases, provides a benchmark focused specifically on occlusion-heavy ball tracking.
Reading between the lines
- Editorial extension: the same temporal occlusion-handling mechanism should generalize to other small, fast objects such as shuttlecocks, cricket balls, or pucks, provided training clips are long enough for the network to learn each object's bounce dynamics.
- Editorial extension: if the reported RMSE drop holds, single-frame appearance-based detectors are probably at a structural disadvantage under occlusion; future gains are more likely to come from trajectory priors and smoothing than from stronger per-frame features.
- Editorial extension: the visibility-weighted loss could be repurposed for semi-supervised tracking, where only a subset of frames in each clip carries ball labels, potentially reducing annotation cost.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript proposes TOTNet, a temporal occlusion tracking network for ball tracking in sports videos, combining 3D convolutions, visibility-weighted loss, and occlusion augmentation. It also introduces TTA, an occlusion-rich table tennis dataset of 9,159 samples with 1,996 occlusion cases collected from Paralympic matches. The abstract reports large performance improvements over prior state-of-the-art, reducing RMSE from 37.30 to 7.19 and increasing fully-occluded-frame accuracy from 0.63 to 0.80 across four datasets in tennis, badminton, and table tennis. The supplied full text, however, is largely illegible (character-encoding corruption), so the architecture details, equations, dataset splits, and experimental protocol cannot be inspected. This report therefore assesses the abstract-level claims and the completeness of the evidence.
Significance. If the reported results are reproducible, TOTNet would be a practically valuable contribution to sports-video ball tracking under occlusion, and the TTA dataset plus the announced code/data release would help the community. The collaboration with Paralympics Australia gives the work a real-world grounding. However, the significance is conditional: the headline numbers are not interpretable without a verifiable evaluation protocol. As submitted, the evidence for the central claim is incomplete because the baselines, splits, occlusion definitions, and post-processing are not available for inspection.
major comments (4)
- [Abstract (comparison protocol)] The central claim rests on comparing TOTNet with prior state-of-the-art methods, yet no protocol is given: train/test splits, evaluation metrics, baseline versions, hyperparameters, or post-processing. Without evidence that all methods were evaluated under identical conditions, the reported RMSE drop from 37.30 to 7.19 and the fully-occluded accuracy rise from 0.63 to 0.80 cannot be attributed to the proposed method. This is load-bearing; please provide a full experimental setup, per-dataset results, and baseline configurations.
- [Abstract (TTA dataset and evaluation)] The paper introduces TTA with 9,159 samples and 1,996 occlusion cases but does not state how samples are split, how occlusion is defined or annotated, or how TTA is used relative to the other three datasets. If TTA is used for training or evaluation, a leakage/overlap check and class distribution should be reported. Without this, the dataset's contribution and the comparability of the reported fully-occluded accuracy cannot be assessed.
- [Full text / Methods] The body of the submitted manuscript is largely unreadable (garbled character encoding), so the 3D convolution design, the visibility-weighted loss, the occlusion-augmentation procedure, and the training details cannot be inspected. This is not a merely stylistic issue: the central derivation and the implementation needed to reproduce the claims are inaccessible. A clean, fully rendered version of the manuscript is required.
- [Evaluation / statistical validity] No error bars, multiple-seed results, or statistical significance tests are reported. For a performance improvement of the reported magnitude, the paper should show variance across runs and ablations. Without ablations isolating the 3D convolution, the visibility-weighted loss, and the occlusion augmentation, it is unknown which component drives the gain or whether the gain is stable.
minor comments (3)
- [Abstract] Typo: 'TOTNets effectiveness' should be 'TOTNet's effectiveness'.
- [Abstract / Code release] The code/data statement mentions a GitHub link but no license; please specify the license and what exactly is released (dataset annotations, pretrained weights, evaluation scripts).
- [Abstract / Scope] The abstract says the method is for 'offline sports analytics' but also emphasizes 'fast-paced scenarios'; please clarify the runtime assumptions and whether the method is intended for offline use only.
Circularity Check
No circularity identifiable: the abstract reports an empirical architecture and external benchmark comparisons, and the garbled full text provides no decipherable derivation that could reduce to its inputs.
full rationale
The available material is the abstract plus a heavily corrupted (mojibake) full text that contains no decipherable equations, derivations, or train/test protocol. The paper's claims are empirical: TOTNet reduces RMSE from 37.30 to 7.19 and improves fully-occluded-frame accuracy from 0.63 to 0.80 on external datasets across tennis, badminton, and table tennis. These are comparisons against prior state-of-the-art methods on independent benchmarks, so the central result is externally falsifiable rather than forced by construction. No fitted parameter is renamed as a prediction in the visible text, no definition of a key quantity is given in terms of the target result, and no load-bearing self-citation is present. The reader's concern about protocol comparability is a correctness/verifiability risk, not a circularity risk: even if splits or occlusion definitions differed, that would make the comparison unfair or uninterpretable, but it would not make TOTNet's output logically identical to its inputs. Because the concrete reduction required by the circularity standard cannot be exhibited from the provided text, the appropriate finding is no significant circularity.
Assumptions & free parameters
Cite this review
Pith. "Pith review of TOTNet: Occlusion-Aware Temporal Tracking for Robust Ball Detection in Sports Videos." pith.science (2026). https://pith.science/paper/DUWKSSCB
@misc{pith2026250809650,
author = {Pith},
title = {Pith review of: TOTNet: Occlusion-Aware Temporal Tracking for Robust Ball Detection in Sports Videos},
year = {2026},
howpublished = {\url{https://pith.science/paper/DUWKSSCB}},
note = {Machine review of arXiv:2508.09650}
}
read the original abstract
Robust ball tracking under occlusion remains a key challenge in sports video analysis, affecting tasks like event detection and officiating. We present TOTNet, a Temporal Occlusion Tracking Network that leverages 3D convolutions, visibility-weighted loss, and occlusion augmentation to improve performance under partial and full occlusions. Developed in collaboration with Paralympics Australia, TOTNet is designed for real-world sports analytics. We introduce TTA, a new occlusion-rich table tennis dataset collected from professional-level Paralympic matches, comprising 9,159 samples with 1,996 occlusion cases. Evaluated on four datasets across tennis, badminton, and table tennis, TOTNet significantly outperforms prior state-of-the-art methods, reducing RMSE from 37.30 to 7.19 and improving accuracy on fully occluded frames from 0.63 to 0.80. These results demonstrate TOTNets effectiveness for offline sports analytics in fast-paced scenarios. Code and data access:\href{https://github.com/AugustRushG/TOTNet}{AugustRushG/TOTNet}.
Forward citations
Cited by 1 Pith paper
-
Demystifying the Role of Rule-based Detection in AI Systems for Windows Malware Detection
Training a malware-detection ML model solely on signature-negative samples improves robustness to adversarial examples and temporal drift at the cost of a false-positive floor.
Reviewed August 5, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.