{"id":"e6e48843-f601-4522-9c56-3c58182ff436","arxiv_id":"2505.05722","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"SurgTracker adapts synthetic-trained point trackers to surgical video with self-distillation and cycle-consistent pseudo-label filtering, improving STIR accuracy over CoTracker3 without any manual annotations.","lead":"SurgTracker adapts a synthetic-trained point tracker to surgical videos using unlabeled data: a frozen teacher generates pseudo-labels, and a cycle-consistency filter keeps only temporally consistent tracks before fine-tuning the student. A generalist reader might care because it shows a simple self-distillation recipe can beat multi-teacher ensembles in a high domain-shift setting.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 0.74 MEE gain over CoTracker3 is not yet established: α is selected on the STIR test set, ~20 test sequences are filtered post hoc, and there are no error bars; the cycle-consistency premise is secondary because unfiltered self-distillation also improves.","rationale":"The reader's CONDITIONAL verdict is appropriate, but the stated weakest assumption is not the most load-bearing part of the argument. The cycle-consistency filter is only responsible for a small portion of the reported improvement: Table 2 shows that unfiltered self-distillation already lowers MEE from the teacher's 17.01 to 16.69, so even if the filter is unreliable, the headline gain over CoTracker3 could still be real. The more dangerous failure mode is evaluation bias: α is selected on the STIR test set, roughly 20 STIR sequences are excluded post hoc, and the reported differences are small with no error bars. A paired bootstrap over sequences and a validation-set α selection would settle whether the 0.74 MEE gap is genuine. If the gap survives, the paper's core claim is supported; if not, it should be weakened. This is the same overall verdict as the reader, but for a different primary reason, hence partial agreement.","tokens_in":6741,"tokens_out":7929,"duration_ms":84847,"concrete_test":"Re-run the STIR evaluation with a strict held-out protocol: split the STIR sequences into validation and test sets, select α on validation rather than on test, keep all test sequences (or pre-register exclusion criteria before seeing model outputs), and compute a paired bootstrap 95% confidence interval over the per-sequence MEE difference between SurgTracker and CoTracker3, plus at least three independent training seeds. If the confidence interval includes 0 or the gain disappears under validation-based α selection, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SurgTracker outperforms CoTracker3 on STIR (Table 1: MEE 16.27 vs 17.01) using only unlabeled Cholec80 videos. For this claim to be trustworthy, the evaluation must be a fair, stable comparison. Three elements of the reported protocol undermine this. First, Sec. 4.2 sets the cycle threshold α=5 after inspecting Table 2, an ablation computed on the STIR benchmark; selecting a hyperparameter on the test set inflates the expected gain. Second, Sec. 4.1 states that 'around 20 sequences with excessive label noise' are filtered out of the evaluation; if the exclusion is post hoc and influenced by model outputs, it biases the comparison. Third, the reported gain is only 0.74 pixels MEE (about 4.3% relative), with no error bars, seeds, or code, so it may be within run-to-run noise. The reader's proposed weak point, the reliability of Eq. 1's cycle-consistency filter, is real: a forward and backward track can both lock onto the same wrong but temporally coherent point and pass the endpoint check. However, Table 2 shows that even with no filtering SurgTracker (16.69 MEE) beats CoTracker3 (17.01), so the headline improvement does not stand or fall on the filter. The evaluation protocol is therefore the more load-bearing concern.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SurgTracker, a semi-supervised framework that adapts the synthetic-pretrained point tracker CoTracker3 to surgical videos using only unlabeled Cholec80 videos. Pseudo-labels are generated online by a frozen teacher that is architecturally identical and identically initialized to the student, and are filtered by a cycle-consistency check (Eq. 1) that discards trajectories whose forward-backward endpoint error exceeds a threshold α. The student is then fine-tuned with a Huber tracking loss. Experiments on the STIR benchmark report that SurgTracker outperforms CoTracker3 and other baselines on MEE, MCD, and <δx_avg, with MEE improving from 17.01 to 16.27. Ablations show that cycle-consistency filtering helps and that a single teacher outperforms multi-teacher supervision in this domain.","tokens_in":7034,"tokens_out":3144,"duration_ms":35629,"significance":"If the reported results hold, the paper makes a useful empirical contribution: it demonstrates that self-distillation with a single architecture-aligned teacher plus cycle-consistency filtering can adapt a synthetic-pretrained point tracker to a high-shift, low-annotation domain such as surgery. The work is also relevant to the broader semi-supervised point-tracking literature because it challenges the assumption that teacher diversity is necessary for pseudo-label quality, and it provides an ablation (Table 3) showing that a single teacher outperforms multi-teacher configurations on this benchmark. The method is simple and computationally efficient, and the unfiltered self-distillation result (16.69 MEE) still beats the teacher baseline (17.01 MEE), which suggests the core idea is not entirely dependent on the filtering threshold. However, the evaluation protocol currently weakens confidence in the exact reported numbers: the cycle threshold is selected on the STIR test set, approximately 20 test sequences are excluded post hoc, and no error bars or multiple-seed results are provided. The paper does not release code or checkpoints, which limits reproducibility.","major_comments":[{"comment":"The cycle-consistency threshold α is selected using the STIR benchmark itself: §4.2 states α=5 'provides the best trade-off,' and Table 2 is an ablation computed on STIR. Because the reported headline gain over CoTracker3 is only 0.74 pixels MEE, tuning α on the test set can materially inflate this gain. The authors should either select α on a held-out validation split of STIR (or on a separate validation set) and report the corresponding test numbers, or clearly state that all Table 1/2 numbers are exploratory and provide a corrected comparison with α fixed before seeing STIR.","section":"§4.2, §4.4, Table 2"},{"comment":"The evaluation excludes 'around 20 sequences with excessive label noise' from the STIR benchmark, but the manuscript does not specify the exact number, the criterion used to identify these sequences, or whether the exclusion was decided before or after inspecting model predictions. Post hoc test-set filtering can bias the comparison in favor of the proposed method. The authors should report the precise filtering rule, list or characterize the excluded sequences, and provide results with and without the exclusion to show robustness.","section":"§4.1"},{"comment":"No error bars, confidence intervals, or multiple-seed experiments are reported for any of the tables. The reported MEE improvement over CoTracker3 is 0.74 pixels (about 4.3% relative), which may be within run-to-run or sequence-level noise given that the training procedure involves random sequence sampling, random strides, and stochastic optimization. The authors should report standard deviations over at least three seeds or bootstrap confidence intervals over test sequences, and ideally release code or checkpoints so the community can verify the comparison.","section":"§4.3, Table 1"},{"comment":"The cycle-consistency filter checks only the Euclidean distance between the original query point and the endpoint of the backward track. This is a weak proxy for pseudo-label quality: a forward trajectory and a backward trajectory can both lock onto the same wrong but temporally coherent point and still pass the endpoint check, and reverse tracking in the presence of occlusion may produce unreliable endpoints. Table 2 shows that even without filtering SurgTracker improves over CoTracker3, so the headline claim does not rest solely on this filter, but the paper's central design rationale does. The authors should analyze filter quality (e.g., precision of kept versus discarded pseudo-labels on a labeled validation set, or a study of failure cases) to substantiate the claim that cycle consistency selects high-quality trajectories.","section":"§3.2.2, Eq. (1)"}],"minor_comments":[{"comment":"There is a spacing artifact in the title ('Y ou Are Y our Best Teacher') and the abstract contains the same issue; this should be corrected.","section":"Title and abstract"},{"comment":"The loss formulation uses a Huber threshold of 6 and an occlusion down-weighting factor of 1/5, but the paper does not state whether these match CoTracker3 or are newly introduced; please clarify the provenance of these hyperparameters.","section":"§3.2.3, Eq. (2)"},{"comment":"The description of the STIR benchmark says it comprises 'around 425' videos and annotations of 'over 3,000 points,' but the exact split sizes and the number of sequences used after filtering are not given; please provide exact counts.","section":"§4.1"},{"comment":"The qualitative figure caption mentions red, green, blue, and pink markers, but the text does not explain what the green dots and pink lines represent; please make the caption self-contained.","section":"Figure 2"},{"comment":"The differences between α=2.5 and α=7.5 are small (MEE 16.76 and 16.43 vs. 16.27 at α=5); the text should discuss whether these differences are meaningful given the absence of uncertainty estimates.","section":"§4.4, Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a simple and potentially useful adaptation method, and the core finding—that self-distillation with a single teacher helps in a high-shift domain—is plausible. However, the evaluation protocol has three interlocking problems (test-set threshold selection, post hoc sequence exclusion, and missing uncertainty quantification) that all affect the headline numerical claim. These are fixable within the scope of a revision, but they need to be addressed before the results can be trusted. I would also encourage the editor to request code or checkpoints, since the reported gains are small and independent verification is otherwise difficult."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a plausible, clearly written empirical study of semi-supervised adaptation of point trackers to surgical video. The headline result — that self-distillation from a single frozen teacher beats a multi-teacher ensemble under domain shift — is a genuinely new empirical finding, and the paper gives reasonable ablation support for it. The exact numbers, however, should be read with caution: the cycle threshold is selected on the STIR test set, about 20 test sequences are filtered post hoc, and there are no error bars.\n\nWhat's new: BootsTAP and CoTracker3 have used teacher-student self-training for point tracking on natural video, but surgery is a high-shift domain where this hasn't been tested. The paper's twist is to use a single teacher identical to the student, rather than a diverse ensemble, and to filter pseudo-labels by cycle consistency. Table 3 shows single-teacher self-distillation does better than multi-teacher configurations on STIR, which is a useful counterpoint to the ensemble approach.\n\nWhat works: the method is simple, the ablations are informative (Tables 2 and 3), and the writing is clear. Importantly, the headline improvement does not depend on the cycle filter: Table 2 shows that even with no filtering, self-distillation (16.69 MEE) beats the CoTracker3 baseline (17.01). The filter adds a small further gain (16.27), but it's a secondary contributor.\n\nWhere the soft spots are: the evaluation protocol undermines the exact numbers. The cycle threshold α is chosen by inspecting the STIR test-set ablation (Table 2), which inflates the expected gain. Section 4.1 says \"around 20 sequences\" with \"excessive label noise\" are filtered out of the 425-sequence STIR set; if this is post hoc, it biases the comparison. The reported gain over CoTracker3 is 0.74 pixels MEE (about 4% relative) with no error bars, seeds, or released code, so it's possible the advantage is within run-to-run noise. These are fixable issues: a proper validation split for hyperparameters, a fixed evaluation set, multiple seeds, and code would make the claim much stronger.\n\nOverall: the central idea is plausible and the study is worth engaging with. It's a reasonable contribution to surgical vision and semi-supervised tracking, not a field-reorienting result. I'd send it to peer review — it deserves referee time — but I'd ask the authors to tighten the evaluation and release code before publication.","headline":"A plausible semi-supervised adaptation study with a genuinely new single-teacher finding, but test-set tuning and missing error bars make the exact gains uncertain.","tokens_in":7542,"tokens_out":2213,"would_cite":false,"duration_ms":20758,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A point tracker can adapt to surgical video by learning from its own filtered output.","keywords":["surgical point tracking","semi-supervised learning","cycle consistency","self-distillation","pseudo-label filtering","domain adaptation","Cholec80","STIR benchmark"],"falsifier":"On the STIR ground-truth annotations, compute for every teacher pseudo-trajectory both the cycle error of Eq. (1) and the true endpoint error against ground truth, and check whether trajectories with cycle error below $\\alpha = 5$ have systematically smaller true error than those above it. If the correlation is weak, or if many low-cycle-error trajectories are far from ground truth, the filtering premise fails.","tokens_in":6549,"feed_emoji":"🎯","tokens_out":7544,"duration_ms":77148,"temperature":0.7,"pith_summary":"The paper tries to establish that a point tracker trained on synthetic data can be adapted to surgical video without any manual annotations, using only 80 unlabeled videos and a single frozen copy of the model as its own teacher. The proposed framework, SurgTracker, generates pseudo-trajectories on unlabeled surgical footage and keeps only the ones that survive a forward-then-backward tracking check, then fine-tunes the student on those filtered labels. If the claim holds, it matters because manual point-level annotation in surgery is costly and impractical, and methods that work on natural video—teacher ensembles or heavy augmentation—have not been shown to survive the larger domain shift of surgical scenes. On the STIR benchmark the paper reports that SurgTracker lowers mean endpoint error from 17.01 to 16.27 relative to its teacher and baseline, CoTracker3, and beats all compared trackers.","feed_headline":"A tracker's own filtered output is its best teacher on surgical video","feed_subtitle":"Fine-tuned on just 80 unlabeled surgical videos with a cycle check, it beats every baseline on the STIR benchmark.","key_machinery":"The central mechanism is the cycle-consistency filter defined by Eq. (1), paired with a fixed teacher identical to the student. For each query point, the frozen teacher tracks forward to the end of the sequence, then tracks backward from the predicted endpoint to the start; the Euclidean distance between the original query point and the backward-tracked endpoint is the cycle error. Trajectories whose error is below $\\alpha = 5$ are kept as pseudo-labels for student fine-tuning. Because the teacher never updates, the supervision signal is stable across training batches, and because teacher and student share architecture and initialization, the student is asked to imitate a representation it already contains. The threshold $\\alpha$ is the main dial controlling the trade-off between label quantity and label quality.","core_discovery":"On its own terms, the paper's discovery is that filtered self-distillation from a single architecture-aligned teacher is enough to adapt a synthetic-pretrained point tracker to a high-shift domain. SurgTracker initializes both teacher and student from CoTracker3 (Online), freezes the teacher, and uses it to predict trajectories on 80 unlabeled Cholec80 videos. A trajectory is retained only if its cycle error—the Euclidean distance between the original query point and the point recovered by tracking forward and then backward—is below a threshold $\\alpha = 5$. The student is then fine-tuned with a Huber tracking loss that pays more attention to visible points and later refinement iterations. The reported results on STIR are MEE 16.27 versus 17.01, MCD 17.12 versus 17.81, and average accuracy 68.55 versus 68.11 against CoTracker3, with ablations showing that removing the filter hurts and that a single teacher beats multi-teacher ensembles.","pith_inferences":["The paper leaves open whether the same single-teacher cycle-consistent recipe transfers to other high-shift, label-scarce domains such as endoscopy, ultrasound, or microscopy; that is a natural next test.","The reported gains over CoTracker3 are modest (0.74 pixels in mean endpoint error), so the practical value for downstream clinical tasks will depend on whether endpoint errors of this size change decisions; the paper does not test that.","Because query points are sampled with SIFT, supervision is concentrated on textured, trackable tissue; low-texture structures such as smooth organs or specular regions may receive few pseudo-labels, and the paper does not measure per-region accuracy.","The same cycle-consistency principle could be applied at test time to re-score or repair trajectories, not just to filter training labels; the paper only uses it during training."],"forward_implications":["If the central claim is correct, surgical point tracking can be improved with unlabeled footage alone, removing the annotation bottleneck that currently limits tissue and tool tracking.","A single consistent teacher can outperform teacher ensembles under large domain shift, so the extra cost and complexity of multi-teacher pseudo-labeling may not be justified in high-shift domains.","Cycle-consistency filtering is doing real work: the ablations show that unfiltered pseudo-labels degrade accuracy on the STIR benchmark.","Because the teacher is frozen and only one model is kept in memory, the adaptation pipeline is cheaper than ensemble-based alternatives, which matters for real-time clinical deployment."],"supporting_citations":[{"why":"Supplies the teacher model, the student initialization, the tracking loss formulation, and the main baseline (CoTracker3 Online).","marker":"[8]"},{"why":"Provides the STIR benchmark with annotated surgical trajectories and the SurgMotion domain-adaptation baseline.","marker":"[18]"},{"why":"Supplies the Cholec80 dataset, the 80 unlabeled laparoscopic videos used for semi-supervised training.","marker":"[16]"},{"why":"Defines the BootsTAP pseudo-labeling baseline that SurgTracker compares against and extends to a high-shift domain.","marker":"[5]"},{"why":"Defines the TAP-Vid metrics (MEE, MCD, average accuracy) used for evaluation.","marker":"[3]"},{"why":"Provides SIFT keypoints used to sample trackable query points from the first frame.","marker":"[11]"},{"why":"Serves as a baseline point tracker (TAPIR) in the STIR comparison.","marker":"[4]"},{"why":"Serves as the classical optical-flow baseline (RAFT) in the STIR comparison.","marker":"[15]"}],"fun_headline_variants":["Self-distillation with cycle check adapts trackers to surgery","One frozen teacher beats ensembles for surgical point tracking","Filtered self-teaching tunes trackers on just 80 surgical videos","Cycle-consistent self-distillation: a single teacher suffices","SurgTracker: self-taught point tracking for high-shift surgery"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the frozen teacher's own cycle error—how close a point comes back to where it started after tracking forward and then backward—tells which forward trajectories are trustworthy; if the teacher fails symmetrically or occlusions break the round trip in misleading ways, the filter will keep bad pseudo-labels or throw away good ones.","fun_headline_variants_meta":{"raw":{"variants":["Self-distillation with cycle check adapts trackers to surgery","One frozen teacher beats ensembles for surgical point tracking","Filtered self-teaching tunes trackers on just 80 surgical videos","Cycle-consistent self-distillation: a single teacher suffices","SurgTracker: self-taught point tracking for high-shift surgery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000481,"raw_usage":{"total_tokens":2384,"prompt_tokens":954,"completion_tokens":1430,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":1343}},"tokens_in":570,"tokens_out":1430,"duration_ms":13879,"temperature":1.0,"reasoning_tokens":1343,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:57:46.393528+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the STIR ground-truth annotations, compute for every teacher pseudo-trajectory both the cycle error of Eq. (1) and the true endpoint error against ground truth, and check whether trajectories with cycle error below $\\alpha = 5$ have systematically smaller true error than those above it. If the correlation is weak, or if many low-cycle-error trajectories are far from ground truth, the filtering premise fails.","supporting_citations":[{"cited_title":"Tracking Everything in Robotic-Assisted Surgery","cited_arxiv_id":"2409.19821","evidence_quote":"Provides the STIR benchmark with annotated surgical trajectories and the SurgMotion domain-adaptation baseline."},{"cited_title":"En- donet: a deep architecture for recognition tasks on laparo- scopic videos","cited_arxiv_id":null,"evidence_quote":"Supplies the Cholec80 dataset, the 80 unlabeled laparoscopic videos used for semi-supervised training."},{"cited_title":"Bootstap: Bootstrapped training for tracking-any-point","cited_arxiv_id":null,"evidence_quote":"Defines the BootsTAP pseudo-labeling baseline that SurgTracker compares against and extends to a high-shift domain."},{"cited_title":"Object recognition from local scale-invariant features","cited_arxiv_id":null,"evidence_quote":"Provides SIFT keypoints used to sample trackable query points from the first frame."},{"cited_title":"Tapir: Tracking any point with per-frame initialization and temporal refinement","cited_arxiv_id":null,"evidence_quote":"Serves as a baseline point tracker (TAPIR) in the STIR comparison."}],"review_version":1}