{"id":"85ca6129-8236-4e24-b46e-e07cf2d89fb7","arxiv_id":"2504.18756","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A boundary-aware multi-stage transformer-TCN with hierarchical sliding-window attention reports improved F1@25/50 action segmentation on SAR-RARP50, Cholec80, and JIGSAWS, but with mixed gains on accuracy and edit score.","lead":"The paper introduces MSBATN, a multi-stage transformer-TCN model that refines action segmentation in surgical videos by jointly predicting actions and segment boundaries with a Gaussian-weighted loss. The reported gains on boundary-localization metrics are modest and mixed across datasets, and the state-of-the-art claim is only partially supported by the paper's own tables.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The F1@25/50 SOTA claim is not anchored to a stated protocol: the configuration named as theirs in Table 3 outperforms the final model in Table 2 on the headline metrics, and no split attribution is given for the ablation.","rationale":"The reader's weakest assumption (the boundary target b_t is undefined) is real, but the empirical claim is more directly endangered by the Table 2 versus Table 3 discrepancy. A SOTA claim is only as good as its test protocol, and the paper leaves the protocol implicit while the ablation table's best configuration does not correspond to the final row. I read this as a possible validation-versus-test mismatch rather than deliberate misreporting, but the absence of split labels and code means the reader cannot tell. The difference is large enough to affect the headline, so conditional acceptance with released code, exact configuration, and a stated split is appropriate, and the verdict should remain conditional. I partially agree with the reader because we converge on reproducibility, but I would rank the split/protocol mismatch ahead of the boundary-target definition as the primary blocker.","tokens_in":17468,"tokens_out":11340,"duration_ms":105663,"concrete_test":"Obtain the code and exact configuration for the ACHDSWO variant, including the definition of b_t in Eq. 5 and the train/validation/test split, then re-run Table 2 on the official test splits of SAR-RARP50, Cholec80, and JIGSAWS. If the reproduced F1@25/50 values match Table 2 within the reported standard deviations, the concern is resolved. If they match Table 3's higher ACHDSWO values instead, the final model and the ablation configuration are inconsistent, and the headlined SOTA comparison must be recomputed with one declared protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MSBATN achieves superior F1@25/50 (Table 2: SAR-RARP 72.4/63.8, Cholec80 73.1/60.2, JIGSAWS 76.1/70.7). The paper does not state whether the ablation in Table 3 is computed on the same test split that produced Table 2. Section 4.3.1 identifies ACHDSWO as the proposed configuration, yet Table 3's ACHDSWO row reports F1@25/50 of 75.0/64.0 on SAR-RARP and 78.3/68.6 on Cholec80, versus 72.4/63.8 and 73.1/60.2 for the final MSBATN in Table 2; on JIGSAWS the ACHFSWO row reports 79.6/72.3 versus 76.1/70.7. If the ablation is on a validation fold, Table 2 can still be a legitimate test result, but the paper must say so and specify the split. If the ablation is on the test set, the final model is not the best configuration and the SOTA comparison is not reproducible. Either way, the headline F1 values are not anchored to a single declared protocol. The under-specified boundary target b_t in Eq. 5 is a separate re-implementation blocker, but the protocol ambiguity directly affects whether the reported SOTA numbers are meaningful.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MSBATN, a multi-stage TCN-Transformer architecture for surgical action segmentation in untrimmed videos. The model uses DINOv2 features, an acausal TCN encoder, a transformer encoder with hierarchical sliding window attention (HSWA), and three TCN decoders. The training loss is a weighted combination of focal, Dice, Gaussian-weighted cosine similarity, and a truncated MSE boundary loss. Experiments are reported on SAR-RARP50, Cholec80, and JIGSAWS, and the abstract claims state-of-the-art F1@25/50 scores. The paper also presents ablations over attention configurations, a comparison of the proposed loss versus binary cross-entropy, and a computational-cost table.","tokens_in":17825,"tokens_out":6483,"duration_ms":55137,"significance":"If the method were fully specified and the reported numbers were tied to a clear evaluation protocol, the boundary-aware weighting idea and the multi-stage refinement architecture could be a useful contribution to the surgical video analysis community. The proposed Gaussian-weighted cosine-similarity loss is a plausible alternative to binary boundary detection and may help with over- and under-segmentation. However, the current evidence is weakened by an undefined boundary target, an unexplained mismatch between the ablation table and the final results, and the use of non-standard data splits. These issues must be resolved before the work can be considered a reliable state-of-the-art result.","major_comments":[{"comment":"The boundary target b_t in Eq. (5) is never formally defined. The text mentions a 5% transition buffer zone at the start and end of each segment, but it does not specify whether b_t is a binary mask, a distance transform, or a probability, nor does it give the Gaussian width sigma and truncation threshold tau used in G_boundary(t) and the min operation. Without this definition, the central boundary-aware loss cannot be reconstructed, and the claimed boundary-refinement mechanism is not reproducible.","section":"Section 3.3, Eq. (5)"},{"comment":"The paper identifies ACHDSWO as the proposed configuration in Section 4.3, but the ACHDSWO row in Table 3 reports F1@25/50 of 75.0/64.0 on SAR-RARP and 78.3/68.6 on Cholec80, whereas the final MSBATN row in Table 2 reports 72.4/63.8 and 73.1/60.2 on the same datasets. The manuscript does not state whether the ablation in Table 3 is computed on the same test split as Table 2. If it is a validation split, that should be stated; if it is the test split, the final model is not the best configuration. Either way, the headline F1 values are not anchored to a single declared protocol.","section":"Tables 2 and 3, Section 4.3"},{"comment":"The paper states 'We split the dataset into 80-20 ratios for training testing for Jigsaw and Cholec80.' For JIGSAWS, the standard evaluation protocol is leave-one-user-out; a random 80-20 split is not comparable to published numbers. No split is described for SAR-RARP50. If Table 2 uses different data partitions than the prior work it compares against, the SOTA comparison is invalid. The exact experimental protocol, including split definitions and any validation-set usage, must be reported.","section":"Section 4, dataset splits"},{"comment":"The claim of state-of-the-art performance is not consistently supported by Table 2. On Cholec80, the F1@50 of 60.2 +/- 8.4 overlaps with ASRF's 60.1 +/- 3.4 within one standard deviation. On JIGSAWS, the accuracy of 86.1 +/- 6.3 is substantially below ASFormer's 92.8 +/- 3.3. The authors should either provide statistical significance tests or temper the abstract's claim of 'state-of-the-art performance' to reflect the mixed results.","section":"Table 2, Cholec80 and JIGSAWS rows"},{"comment":"The aggregated attention formula in Eq. (1) uses undefined quantities: the attention scores e^s_ij, the neighborhood sets N_i^s, and the scale weights w_s are not formally defined, and it is unclear whether w_s is learned or fixed in the implementation. This makes the HSWA module description incomplete and hinders reproducibility.","section":"Section 3.2.2, Eq. (1)"}],"minor_comments":[{"comment":"The header line in Table 2 is malformed ('F1@50SAR-RARP' should be split), and the table layout should be fixed.","section":"Table 2 header"},{"comment":"The MS-TCN row cites Li et al. (2020), but MS-TCN is introduced in Farha and Gall (2019); the reference list entry Li et al. (2020) is for MS-TCN++.","section":"Table 2, MS-TCN row"},{"comment":"The sentence 'considering this 5% of the total segment length is labeled as a transition buffer zone' is grammatically incomplete and should be rewritten to clearly define the buffer-zone rule.","section":"Section 3"},{"comment":"Table 4 does not specify the dataset used for the loss comparison; this should be stated in the caption.","section":"Table 4"},{"comment":"The heading 'Feature Work' is a typo for 'Future Work'.","section":"Conclusion"},{"comment":"The font size in Figure 3 is very small, and some component labels are difficult to read; the figure should be enlarged or redrawn.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The most pressing issue is the inconsistency between Table 2 and Table 3: the ablation configuration named as the proposed one outperforms the final model on the headline metrics, and no split attribution is given. If the authors cannot clarify this protocol, the SOTA claim is not credible. The undefined boundary target b_t is also a blocker for any reproducibility effort. I recommend that the editor require the authors to provide the missing definitions, a clear experimental protocol, and a justification of the chosen data splits before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The MSBATN paper is a reasonable architecture contribution: a multi-stage TCN-transformer that combines sliding-window attention with hierarchical temporal attention for surgical action segmentation, plus a composite loss with Gaussian-weighted similarity and boundary MSE. The specific combination is new, and the ablation studies show that the components matter. The efficiency comparison is a nice touch. The motivation is clear and the design choices are defensible.\n\nThat said, the central evaluation claim has a real protocol problem. Table 2 presents MSBATN as the final model, but Table 3 identifies ACHDSWO as the proposed configuration, and ACHDSWO reports higher F1@25/50 on all three datasets (e.g., 75.0/64.0 vs 72.4/63.8 on SAR-RARP; 78.3/68.6 vs 73.1/60.2 on Cholec80). The paper never states whether Table 3 is computed on a validation split or the test set. If it is validation, the reader needs to know that. If it is test, then the final model is not the best configuration found, and the SOTA comparison in Table 2 is not reproducible from the paper's own numbers. This is not a minor omission; it directly affects whether the headline 'superior F1 scores' claim is anchored to a single declared protocol.\n\nA second blocker is the boundary target b_t in Eq. 5. The paper says it is the ground truth boundary position at frame t, but does not say whether it is a binary mask with the 5% buffer zone, a distance map, or a probability. The Gaussian width sigma and truncation threshold tau are also unspecified. Without these, the boundary-aware loss cannot be reimplemented. The loss weights are given, but the target definition is missing.\n\nMore minor: the gains over prior work are sometimes within standard deviations or inconsistent across metrics. On Cholec80 F1@50, MSBATN is 60.2 vs ASRF's 60.1. On JIGSAWS, accuracy is 86.1 vs ASFormer's 92.8. The paper leans on F1@25/50, which is fair, but 'competitive results across other metrics' is doing a lot of work. Also, no code or feature-extraction details are released, so verification is limited.\n\nOverall, the architecture is sound and the problem is relevant. The paper needs major revision: clarify the split for Table 3, define b_t and the hyperparameters, report results consistently, and release code. I would send it to peer review because the core idea is worth engaging with, but the current version does not support the SOTA claim as written.","headline":"A coherent architecture paper whose headline SOTA claim is undercut by an unspecified protocol: the named configuration in the ablation beats the final model on the same metrics, and the boundary loss is under-specified.","tokens_in":18338,"tokens_out":2000,"would_cite":false,"duration_ms":21357,"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":"MSBATN, a multi-stage TCN-transformer with hierarchical sliding window attention and a Gaussian-weighted boundary-aware loss, reports the highest F1@25 and F1@50 scores on SAR-RARP50, Cholec80, and JIGSAWS surgical action segmentation.","keywords":["action segmentation","surgical video analysis","boundary detection","transformer","temporal convolutional network","sliding window attention","boundary-aware loss","surgical workflow analysis"],"falsifier":"Reproduce MSBATN on SAR-RARP50 under two plausible definitions of the boundary supervision $b_t$ (a binary mask with the 5% buffer zone described in Section 3, or a distance-to-boundary map), keeping the reported hyperparameters; if neither reproduces F1@25=72.4 and F1@50=63.8 within the reported standard deviations across random seeds, the boundary-refinement mechanism as specified cannot explain the claimed improvement.","tokens_in":17246,"feed_emoji":"🏥","tokens_out":10606,"duration_ms":86400,"temperature":0.7,"pith_summary":"Action segmentation in surgical videos is hard because actions vary in duration and their start and end boundaries are ambiguous, which makes frame-wise classifiers over- or under-segment. This paper argues that the fix is to model boundaries explicitly: it proposes MSBATN, a multi-stage network combining dilated temporal convolutions, hierarchical sliding window attention, and a unified loss that weights the interior of action segments and their boundary zones differently. On three surgical datasets, it reports F1@25 and F1@50 scores above previously published methods, with comparable accuracy and edit scores. The sympathetic reader takes away that treating boundary detection as a soft, context-weighted task rather than a binary classification can improve segment-level evaluation of long untrimmed surgical videos.","feed_headline":"New model lifts F1 scores on surgical action boundaries","feed_subtitle":"A boundary-aware transformer beats prior methods on three surgical video datasets.","key_machinery":"The load-bearing object is the boundary-aware segmentation loss, a weighted sum of focal classification loss, Dice loss, a Gaussian-weighted cosine-similarity term that peaks at the temporal centre of each ground-truth segment, and a truncated MSE term that is Gaussian-weighted to emphasize the annotated start and end frames. The complementary architectural mechanism is Hierarchical Sliding Window Attention (HSWA), which keeps one sliding window expanding while another shrinks, and aggregates attention scores across temporally downsampled scales, so the transformer maintains linear cost while covering long sequences. Together they implement the paper's thesis: boundaries are soft and context-dependent, so the loss should weight segment interiors and edges differently, and the attention should resolve ambiguities between local and global temporal structure.","core_discovery":"The central claim is that explicit, soft boundary modelling improves surgical action segmentation more than the binary boundary classifiers used in prior work. MSBATN couples a TCN encoder-decoder with a transformer whose Hierarchical Sliding Window Attention (HSWA) combines a dual sliding window (one expanding, one shrinking across layers, with dilation) and multi-scale hierarchical attention, so the model sees both fine local detail and long-range structure. Its unified loss is the real novelty: a focal term for class balance, a Dice term for segment overlap, a Gaussian-weighted cosine-similarity term that enforces feature smoothness at the core of each action while allowing change at the edges, and a truncated Gaussian-weighted MSE term that concentrates error on the annotated start and end frames. With three refinement decoders operating only on the predicted segments, the model corrects over- and under-segmentation without reprocessing the video. Table 2 reports F1@25=72.4 and F1@50=63.8 on SAR-RARP50, F1@25=73.1 and F1@50=60.2 on Cholec80, and F1@25=76.1 and F1@50=70.7 on JIGSAWS, which the paper frames as superior to ASRF, Asformer, NETE, and the MS-TCN family baselines on segment-level metrics.","pith_inferences":["Beyond the paper: because the boundary supervision label is under-specified, a natural next experiment is to test whether a distance-to-boundary regression target or a single-frame peak label changes the reported F1 margins; this would localize which part of the loss produces the improvement.","Beyond the paper: the Gaussian centre-weighting of cosine similarity may generalize beyond surgery to any long-video action segmentation with ambiguous transitions; swapping the loss into a standard multi-stage temporal network on non-surgical benchmarks would test that transfer.","Beyond the paper: the soft-boundary thesis implies that binary boundary labels are an inferior training signal; replacing the truncated MSE boundary term with a soft distance-map regression and measuring over-segmentation rates would be a direct testable extension."],"forward_implications":["On SAR-RARP50, Cholec80, and JIGSAWS, the model reports F1@25 and F1@50 scores above the cited baselines, with frame accuracy and edit score at par, so the improvement is specifically in segment-level boundary fidelity.","The ablation study in Table 4 shows the Gaussian start-end loss beats binary cross-entropy across causal and acausal settings, implying the boundary-aware loss alone accounts for a substantial part of the reported improvement.","The three-decoder refinement architecture operates only on predicted segments rather than re-processing the video, so the boundary correction adds negligible computational cost, consistent with the reported GFLOPs comparable to Asformer.","The best-performing configuration in the ablation (ACHDSWO) combines acausal convolutions, hierarchical dual sliding windows, and overlap, indicating that the architecture and the loss are separable contributions that each contribute to the reported gains.","The paper's comparisons position MSBATN against ASRF and Asformer as boundary-aware baselines, so its claims are specifically about beating soft-boundary and attention-based refinement, not just frame-level classifiers."],"supporting_citations":[{"why":"Defines the MS-TCN baseline whose over- and under-segmentation problems motivate the need for boundary-aware refinement.","marker":"Farha and Gall (2019)"},{"why":"Extends MS-TCN to MS-TCN++, the dilated multi-stage temporal convolutional family that the MSBATN encoder and decoders build on.","marker":"Li et al. (2020)"},{"why":"Introduces ASRF, the boundary-regression refinement method that is the main boundary-aware baseline MSBATN claims to beat on F1@25 and F1@50.","marker":"Ishikawa et al. (2021)"},{"why":"Asformer, the transformer action-segmentation architecture whose decoder stack and attention design MSBATN adapts and extends.","marker":"Yi et al. (2021)"},{"why":"NETE, an online multi-stage phase-recognition model used as a comparison baseline across all three datasets.","marker":"Yi et al. (2022)"},{"why":"DINOv2-Reg, the self-supervised feature extractor with register tokens that supplies the per-frame spatial features to the temporal model.","marker":"Oquab et al. (2023)"},{"why":"Longformer, the sliding-window sparse attention pattern that inspires the DSWA component of HSWA.","marker":"Beltagy, Peters and Cohan (2020)"},{"why":"SAR-RARP50, the suturing-action benchmark where the best F1@25 and F1@50 gains are reported.","marker":"Psychogyios et al. (2023)"},{"why":"Cholec80, the laparoscopic phase-recognition dataset used as a second benchmark.","marker":"Twinanda et al. (2016)"},{"why":"JIGSAWS, the surgical gesture dataset used as a third benchmark.","marker":"Gao et al. (2014)"}],"fun_headline_variants":["MSBATN tightens action boundaries in surgical videos","Soft boundary loss improves surgical action segmentation","Sliding window attention refines surgical action boundaries","Unified loss lifts F1 in surgical video segmentation","Hierarchical attention reduces over- and under-segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The boundary loss rests on a ground-truth boundary label $b_t$ that the paper never defines: it does not specify whether it is a binary mask, a distance map, or a probability, nor the values of the Gaussian width $\\sigma$ and the truncation threshold $\\tau$, so the central refinement mechanism cannot be reconstructed from the text.","fun_headline_variants_meta":{"raw":{"variants":["MSBATN tightens action boundaries in surgical videos","Soft boundary loss improves surgical action segmentation","Sliding window attention refines surgical action boundaries","Unified loss lifts F1 in surgical video segmentation","Hierarchical attention reduces over- and under-segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000273,"raw_usage":{"total_tokens":1684,"prompt_tokens":1040,"completion_tokens":644,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":570}},"tokens_in":656,"tokens_out":644,"duration_ms":5917,"temperature":1.0,"reasoning_tokens":570,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:09:41.052605+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce MSBATN on SAR-RARP50 under two plausible definitions of the boundary supervision $b_t$ (a binary mask with the 5% buffer zone described in Section 3, or a distance-to-boundary map), keeping the reported hyperparameters; if neither reproduces F1@25=72.4 and F1@50=63.8 within the reported standard deviations across random seeds, the boundary-refinement mechanism as specified cannot explain the claimed improvement.","supporting_citations":[],"review_version":1}