{"id":"e685d522-c6b6-417e-8aa6-706b87500513","arxiv_id":"2507.09577","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"MA-SAM2 adds context-aware and occlusion-resilient memory to SAM2 and reports Challenge IoU of 62.49 percent on EndoVis2017 and 64.40 percent on EndoVis2018, beating SAM2 by 6.10 and 4.36 points.","lead":"A training-free upgrade to SAM2 for surgical video segmentation uses two new memory modules, context-aware and occlusion-resilient, and reports 4.4 to 6.1 percent higher benchmark scores than SAM2 on two surgical datasets. The paper is worth reading if you use SAM2-style zero-shot tracking in long videos with frequent occlusions and want a no-training improvement with only prompt initialization.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains may be selection artifacts of hand-set memory thresholds and small test sets; sensitivity and transfer checks are missing.","rationale":"The reader identified the same weakest assumption: the method's performance depends on several hand-set thresholds and capacities that are neither derived nor validated, and all videos are used for testing. I agree that this is the most load-bearing concern because the central claim is a consistent benchmark improvement, and the evidence consists of point estimates on only 8 and 15 videos. The internal inconsistency in Equation (2), the swapped abstract numbers, and the missing error bars are real but secondary; they do not change the verdict. A threshold-sensitivity sweep plus a cross-dataset transfer evaluation would settle whether the gains are robust or are selection artifacts. Since the reader already returned CONDITIONAL, my stress test does not move the verdict; the conditional status remains appropriate.","tokens_in":7261,"tokens_out":10621,"duration_ms":117952,"concrete_test":"Fix all thresholds using EndoVis2018 alone (grid search over IoU threshold in {0.6, 0.7, 0.8, 0.9}, bounding-box overlap bounds in {[0.5, 0.9], [0.6, 0.9], [0.7, 1.0]}, and ORM capacity in {3, 5, 7}), then evaluate that fixed configuration on EndoVis2017; repeat with the datasets reversed. Additionally, report a per-video paired bootstrap 95% confidence interval for the Challenge IoU difference on both datasets. If the MA-SAM2 advantage over SAM2 disappears under the transfer split or the confidence interval includes zero, the threshold-selection artifact is not ruled out.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section 1 and Tables 1-2 is that MA-SAM2 beats SAM2 by 6.10 and 4.36 Challenge IoU points, training-free and parameter-free. What must be true for that claim to generalize is that the hand-set memory hyperparameters in Section 2.2 are not themselves responsible for the advantage. The paper gives no derivation, no grid search, no sensitivity analysis, and no held-out validation for the IoU threshold of at least 0.8, the bounding-box overlap range of 0.6 to 0.9, the ORM capacity of 5 frames, or the unnamed CAM confidence/IoU thresholds and CAM capacity. Because all 8 or 15 test videos are used in every reported number and no per-video variance or significance testing is provided, the improvement could result from selecting thresholds on the exact test split. With datasets this small, a few tuned choices can shift Challenge IoU by more than the reported 4 to 6 points, so without sensitivity and transfer evidence the consistency of the gains is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MA-SAM2, a training-free modification of SAM2 for surgical video segmentation. It replaces SAM2's greedy FIFO memory with two complementary memory modules: a context-aware memory (CAM) that stores historical high-quality masks through cumulative-score hypothesis selection, and an occlusion-resilient memory (ORM) that filters contaminated masks using IoU and bounding-box overlap criteria. The method is evaluated on EndoVis2017 and EndoVis2018 under zero-shot settings, reporting Challenge IoU improvements of 6.10 and 4.36 points over SAM2, with an ablation showing that ORM provides most of the gain. The code is publicly released.","tokens_in":7435,"tokens_out":4985,"duration_ms":54202,"significance":"If the reported results hold, MA-SAM2 offers a genuinely simple, training-free improvement over SAM2 on two public surgical benchmarks, which is practically useful and reproducible because no learned parameters are introduced and the code is available. The ablation study is a strength: it clearly attributes the main gain to the ORM module rather than to the baseline alone. The significance is, however, moderated by the small test sets and by the fact that several hand-set hyperparameters are neither derived nor tested for sensitivity, so the consistency and generalizability of the gains are not yet established.","major_comments":[{"comment":"The reported gains of 6.10 and 4.36 Challenge IoU points rest on a set of hand-set hyperparameters that are neither derived nor validated: the IoU confidence threshold θ_IoU ≥ 0.8, the bounding-box overlap range 0.6–0.9, the ORM maximum capacity of 5 frames, the unnamed CAM confidence and IoU thresholds, the unnamed CAM maximum capacity, and the constant ε in Eq. (3). Because all 8 EndoVis2017 and 15 EndoVis2018 videos are used for testing and no sensitivity analysis is reported, the improvements in Tables 1 and 2 could result from selecting these thresholds on the exact test split. The paper should report how each threshold was chosen, provide a sweep or sensitivity study around each value, and ideally include a transfer check with thresholds fixed on one dataset and evaluated on the other. The claim of being 'without additional parameters' should also be qualified: no parameters are learned, but several manual hyperparameters are introduced.","section":"§2.2, Tables 1–2"},{"comment":"No measure of variability or statistical significance is provided for any number in Tables 1–3. With only 8 and 15 test videos, the gap between MA-SAM2 and SAM2 could be within video-level variance, and the ablation increments for CAM (0.89 points on EndoVis2018 and 1.47 points on EndoVis2017) are especially small relative to likely video-level noise. The manuscript should report per-video Challenge IoU scores, error bars or bootstrap confidence intervals, and a paired significance test (e.g., Wilcoxon signed-rank or a paired bootstrap) for both the full-model comparison and the ablation rows.","section":"§3, Tables 1–3"},{"comment":"Equation (2) is ambiguous and appears internally inconsistent with the prose: the text says alternative masks are post-processed by removing regions that overlap with Ms and retaining only their largest connected components, but the displayed formula includes the full overlap region Ms ∩ Ma and applies the CC operator only to Ms \\ (Ms ∩ Ma). Since this formula defines the core occlusion-resilient filter, which the ablation shows is the main source of the improvement, it must be unambiguous. Similarly, Eq. (3) writes S(t − 1) on the right-hand side although the left-hand side is S_k(t); the cumulative score should be per-branch, i.e., S_k(t − 1), otherwise the CAM hypothesis-selection rule is not reproducible. Please correct the notation and define the special cases where the overlap region or the non-overlap region is empty.","section":"§2.2, Eqs. (2)–(3)"}],"minor_comments":[{"comment":"The abstract states performance improvements of 4.36% and 6.1% on EndoVis2017 and EndoVis2018, respectively, but Section 1 and Tables 1–2 report 6.10% on EndoVis2017 and 4.36% on EndoVis2018; the order in the abstract should be reversed.","section":"Abstract"},{"comment":"There are several phrasing errors, for example 'a mask-based one-prompt strategy provides a single prompt' reads as a sentence fragment, and 'The SAM2 process each input frame' should be 'SAM2 processes each input frame'.","section":"§1 and §2.1"},{"comment":"The paper emphasizes real-time requirements but reports no runtime or latency comparison against SAM2 or Surgical SAM2; consider adding inference-time measurements or softening the real-time claim.","section":"§1 and §3"}],"recommendation":"major_revision","confidential_remarks":"The central risk is that the reported benchmark gains are selection artifacts of the hand-set memory thresholds. The absence of any sensitivity analysis is the largest gap, and the small test sizes make the statistical question non-trivial. If the authors provide a threshold sweep, per-video results, and a significance test, and if the gains persist, the paper could be acceptable. The abstract inconsistency between the two datasets is also a sign that a careful final proofread is needed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a read if you work on SAM2 memory or surgical tracking. The paper does what it says: it takes SAM2's memory bank, splits candidates into a context-aware branch (cumulative log-IoU scoring, after SAM2Long) and an occlusion-resilient branch (distractor-aware memory, after Videnovic et al., extended to multiple targets), and runs one-prompt multi-target inference. No extra parameters, no training. Reported gains over SAM2 are about 6 points on EndoVis2017 and 4.4 points on EndoVis2018, and the ablation shows most of that comes from the occlusion-resilient memory. The combination is genuinely new, and the one-prompt multi-target loop is a sensible practical contribution.\n\nSoft spots, in order of weight. First, all evaluation is on 8 and 15 videos, with no error bars, no per-video variance, and no significance tests. On datasets this small, a tuned choice on the test split can move Challenge IoU by more than the reported gains. Second, the method's thresholds and capacities are hand-set and not justified: IoU confidence >= 0.8, bbox overlap 0.6-0.9, ORM capacity 5, plus unnamed CAM thresholds and capacity. No sensitivity analysis, no grid search, no held-out validation. The stress-test concern about selection artifacts is legitimate, especially because all videos are used for testing. Third, the abstract has the datasets reversed: it says 4.36% on EndoVis2017 and 6.1% on EndoVis2018, while the tables and Section 1 show the opposite. Fourth, Eq. (2) doesn't match its prose: the text says alternative masks are processed by removing overlap with the primary mask, but the formula keeps the overlap and just applies connected components to the non-overlapping part of Ms, so the result is almost Ms itself. That looks like a typo, but it needs fixing.\n\nThe core claim, that a memory redesign can improve SAM2 without retraining, is supported by the tables and the ablation. The magnitude and generalizability are not fully established because of the tuning and small-test issues, but this is an addressable problem, not a load-bearing flaw. The paper is honest about its scope, the comparison to SAM2Long and distractor-aware memory is fair, and the code is promised.\n\nFor a referee: yes, I would send it out. It is a clean, testable, incremental contribution that belongs in a surgical segmentation venue. The review should ask for error bars, a sensitivity analysis on the thresholds, a correction of the abstract and Eq. (2), and ideally a transfer check to another long-video dataset. If those come back clean, this is a usable result.","headline":"A training-free SAM2 memory redesign that improves surgical instrument tracking on two small benchmarks; the gains are plausible but the hand-set thresholds and lack of error bars leave the size of the effect unproven.","tokens_in":7990,"tokens_out":1861,"would_cite":true,"duration_ms":21133,"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 training-free memory split into context and occlusion branches lifts SAM2's surgical segmentation by 6.1% and 4.36% on two EndoVis benchmarks.","keywords":["surgical instrument segmentation","video object segmentation","SAM2","training-free segmentation","memory-augmented tracking","occlusion handling","context-aware memory","EndoVis benchmarks"],"falsifier":"Sweep the paper's hand-set values on EndoVis2017—IoU threshold from 0.5 to 0.95, bounding-box overlap window, and ORM capacity from 1 to 10—then fix the best values and run them on EndoVis2018 without retuning; if the reported 6.1% and 4.36% gains shrink or reverse, the thresholds were likely tuned to the test videos.","tokens_in":7039,"feed_emoji":"🩺","tokens_out":12645,"duration_ms":112625,"temperature":0.7,"pith_summary":"The paper claims that SAM2's greedy, first-in-first-out memory bank is the main reason it loses track of surgical instruments in long videos, and that a training-free memory redesign addresses this. It introduces MA-SAM2, which separates the memory into a context-aware branch that keeps high-quality historical masks and an occlusion-resilient branch that stores frames with strong interference. With no additional parameters and no training, MA-SAM2 improves Challenge IoU by 6.1 percent on EndoVis2017 and 4.36 percent on EndoVis2018 relative to SAM2. The practical point is that a fix to memory management, rather than a new model, can make foundation-model segmentation usable in the operating room.","feed_headline":"Memory split lifts SAM2 surgical segmentation by 6.1 percent","feed_subtitle":"A training-free memory redesign recovers instruments that SAM2 loses to occlusion and overlap in long surgical videos.","key_machinery":"The load-bearing object is the augmented memory bank, split into two curated stores. The context-aware memory (CAM) carries the best-scoring historical masks selected by a cumulative log-IoU hypothesis test, which combats slow drift; the occlusion-resilient memory (ORM) carries up to five frames flagged as having strong interference, selected by an IoU-confidence threshold and a bounding-box overlap window, which resists sudden occlusion. The bank fills ORM first and CAM only when spare capacity remains, and the outputs of both branches are integrated to build the final segmentation representation.","core_discovery":"The paper's central claim is that SAM2's sequential memory policy pollutes its tracking when intermediate frames are occluded or ambiguous, so the paper replaces that policy with two complementary memories. The context-aware memory scores three candidate masks per frame by a cumulative log-IoU and stores the best historical masks, while the occlusion-resilient memory screens each frame with an IoU-confidence threshold (at least 0.8) and a bounding-box overlap window (0.6–0.9) to identify interference frames, keeping up to five such frames. The two memories are stored in a fixed-capacity bank that prioritizes the occlusion-resilient memory and adds context-aware memory only when capacity remains. In multi-instrument videos, a single initial mask per category is propagated through a one-prompt, multi-target loop. The authors show on EndoVis2017 and EndoVis2018 that this memory split outperforms SAM2 and each memory branch individually, with the occlusion-resilient memory contributing the larger gain.","pith_inferences":["If the hand-set thresholds (IoU at least 0.8, bounding-box overlap in 0.6–0.9, ORM capacity 5, and the unnamed CAM thresholds and capacity) were tuned on the test videos, the benchmark gains could be selection artifacts; the paper does not report how the values were chosen or how sensitive the results are to them.","The same context-plus-occlusion memory split could transfer to other long-video object-tracking domains with frequent occlusion, such as sports or surveillance, since the mechanism does not depend on surgical appearance.","The large gain from ORM suggests that SAM2's practical failures in long videos stem more from its memory update rule than from its image features, which would encourage research on memory-scheduling policies for foundation trackers."],"forward_implications":["On both EndoVis2017 and EndoVis2018, the full MA-SAM2 beats plain SAM2, and the ablation attributes most of the gain to the occlusion-resilient memory rather than the context-aware memory.","The one-prompt, multi-target loop means a video sequence needs only one initial mask per instrument category; no corrective prompts are required during tracking.","Because the method is training-free and adds no parameters, it can be dropped into any existing SAM2-based segmentation pipeline as a memory-policy replacement.","The single-prompt design also speeds up inference relative to per-target processing, which matters for real-time surgical use.","The improvement is concentrated in some instrument categories such as Bipolar Forceps, Monopolar Curved Scissors, and Ultrasound Probe, indicating the memory policy helps most where occlusion and motion are frequent."],"supporting_citations":[{"why":"SAM2 is the base video segmentation model whose greedy FIFO memory MA-SAM2 replaces; it also serves as the main baseline in every comparison table.","marker":"[12]"},{"why":"Supplies the EndoVis2017 surgical video dataset and the Challenge IoU metric used for evaluation.","marker":"[2]"},{"why":"Supplies the EndoVis2018 surgical video dataset used for the second evaluation.","marker":"[1]"},{"why":"Provides the preprocessing protocol the paper follows for the EndoVis2017 videos.","marker":"[15]"},{"why":"Provides the instrument-type segmentation annotations used as ground truth for evaluation.","marker":"[5]"},{"why":"Introduces the cumulative log-IoU hypothesis scoring that the context-aware memory adapts to select stable masks.","marker":"[4]"},{"why":"Introduces the distractor-aware memory idea that the occlusion-resilient memory extends to multi-target inference.","marker":"[17]"},{"why":"Surgical SAM2 is the surgical-specific training-free comparison method that prunes frames; the paper compares against it and contrasts its memory approach.","marker":"[8]"},{"why":"Establishes the zero-shot, instrument-type label-level evaluation protocol the paper adopts for testing all videos.","marker":"[24]"}],"fun_headline_variants":["Memory split boosts SAM2 surgical video segmentation by 6.1%","Training-free two-memory bank improves SAM2 on surgical videos by 6.1%","Occlusion-resilient memory helps SAM2 track instruments in surgery","SAM2's greedy memory replaced: surgical segmentation up 6.1%","Two memories, zero training: SAM2 gains 6.1% on surgery videos"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The improvements hinge on several hand-picked values—a minimum overlap score (IoU) of 0.8, a bounding-box overlap between 0.6 and 0.9, a storage cap of 5 frames, and other unstated thresholds—whose choice is never justified or tested for sensitivity; if those values were tuned on the test videos, the reported gains could be artifacts.","fun_headline_variants_meta":{"raw":{"variants":["Memory split boosts SAM2 surgical video segmentation by 6.1%","Training-free two-memory bank improves SAM2 on surgical videos by 6.1%","Occlusion-resilient memory helps SAM2 track instruments in surgery","SAM2's greedy memory replaced: surgical segmentation up 6.1%","Two memories, zero training: SAM2 gains 6.1% on surgery videos"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001259,"raw_usage":{"total_tokens":5165,"prompt_tokens":960,"completion_tokens":4205,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":4101}},"tokens_in":576,"tokens_out":4205,"duration_ms":33262,"temperature":1.0,"reasoning_tokens":4101,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:52:04.113114+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Sweep the paper's hand-set values on EndoVis2017—IoU threshold from 0.5 to 0.95, bounding-box overlap window, and ORM capacity from 1 to 10—then fix the best values and run them on EndoVis2018 without retuning; if the reported 6.1% and 4.36% gains shrink or reverse, the thresholds were likely tuned to the test videos.","supporting_citations":[{"cited_title":"In: 2018 17th IEEE international conference on machine learning and applications (ICMLA)","cited_arxiv_id":null,"evidence_quote":"Provides the preprocessing protocol the paper follows for the EndoVis2017 videos."},{"cited_title":"In: International Conference on Medical Im- age Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Provides the instrument-type segmentation annotations used as ground truth for evaluation."}],"review_version":1}