{"id":"f5bf2b47-cc3c-4151-b180-22d22bc907de","arxiv_id":"2507.11994","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SAMST refines pseudo-labels with SAM prompts derived from connected regions and threshold filtering, improving semi-supervised remote sensing segmentation on Potsdam at 1/32 labeled data.","lead":"The paper presents SAMST, a semi-supervised method that uses the Segment Anything Model (SAM) to clean and refine pseudo-labels for remote sensing images, then retrains a segmentation network on the improved labels. The method reports higher mean intersection-over-union on the Potsdam dataset when only 1/32 of training patches are labeled, but the evidence has several unclear points.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I is internally hard to reconcile: the supervised baseline has OA 93.48 while every semi-supervised comparison is between 80.57 and 83.83, and SAMST's OA drops to 86.44 even though per-class F1 on dominant classes improves; the mIoU gain may stem from a mismatched evaluation or weak…","rationale":"The reader's weakest assumption focuses on SAM's mask quality and the stitching rule, which is a genuine algorithmic risk. However, the more load-bearing concern is that the reported numbers themselves appear internally inconsistent and may have been produced under non-comparable conditions. If Table I cannot be reproduced from confusion matrices, then the SAM-refinement mechanism is never the issue; the headline improvement itself is unverified. I therefore center the stress test on the quantitative evidence rather than on the SAM assumption. The method is plausible, and the visualizations in Figures 2 and 3 do illustrate boundary corrections and error filtering, so the underlying idea deserves further study. But the paper provides no code, no ablations separating the three modules, no error bars, and no threshold values, and the unexplained OA drop plus the unusually weak comparison methods make the stated state-of-the-art claim premature. A conditional verdict is appropriate: the paper should not be accepted in its current form unless the confusion-matrix check verifies Table I and the comparison methods are re-run under matched training budgets and backbones. If the check fails, the verdict should move to reject.","tokens_in":6600,"tokens_out":7754,"duration_ms":92938,"concrete_test":"Release the test-set confusion matrices for all rows of Table I, or rerun the released code on the same 2016-patch split, and recompute OA, mIoU, and per-class F1 from those matrices. Then check whether an OA drop from 93.48 to 86.44 is compatible with the reported per-class F1 values for Building, Tree, and Surface; if the matrices do not reproduce Table I, or if the confusion matrices are unavailable, the central claim is unsupported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central performance claim rests entirely on Table I, and Table I contains a pattern that is difficult to explain. The supervised baseline trained on 1/32 labeled data reaches OA 93.48, yet every published semi-supervised method compared (LSST 80.57, ClassHyPer 83.62, ST++ 83.83) is 9.65 to 12.91 points lower in OA while also being lower in mIoU. A standard semi-supervised method with the same backbone and training budget should not be uniformly worse than its own supervised baseline on both metrics; this suggests the comparison methods were not run under matched conditions, or that the baseline was evaluated with a different protocol. Within SAMST itself, OA falls from 93.48 to 86.44 (a 7.04-point drop) while mIoU rises from 67.00 to 70.80 and per-class F1 for Building, Tree, and Surface all improve. Those three classes dominate the Potsdam scene, so an OA drop of this size implies a sharp recall decrease on large classes that is neither reported nor discussed. The paper also selects the 'best checkpoint' for pseudo-label generation without stating the selection criterion or a held-out validation split, and the SAM refiner's seven prompt thresholds plus tc/to are 'set through experimental outcomes' with no values or validation protocol. If any of these choices used the test set, the reported numbers are optimistically biased. The weight α of the pseudo-labeled loss and the class weights in Equations (2)-(3) are likewise undisclosed. Consequently, the observed mIoU gain cannot currently be attributed to the SAM-based Pseudo-label Refiner; it could be an artifact of evaluation inconsistency, test-set peeking in threshold or checkpoint selection, or underpowered comparison baselines.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SAMST, a semi-supervised semantic segmentation method for remote sensing images that combines supervised self-training with a SAM-based pseudo-label refiner. The refiner consists of a Threshold Filter Module, a Prompt Generation Module that creates boxes and positive/negative points for SAM, and a Label Refinement Module that stitches SAM masks with model probabilities. A weighted composite loss balances labeled and pseudo-labeled training. The empirical claim is that after a single iteration on the Potsdam dataset with 1/32 labeled data, SAMST improves mIoU from 67.00 (baseline) to 70.80 and outperforms LSST, ST++, and ClassHyPer in mIoU and average F1.","tokens_in":6970,"tokens_out":3562,"duration_ms":39130,"significance":"If fully validated, the idea of using SAM to refine pseudo-labels in semi-supervised remote sensing segmentation would be a timely and useful contribution, and the proposed pipeline is clearly motivated. The method is not circular: pseudo-labels are generated on unlabeled data and the final evaluation is on a held-out test set. However, the current evidence is not sufficient to establish the contribution: the central table has a puzzling internal inconsistency, the refiner depends on several undisclosed hand-set thresholds, no ablations are provided, and comparison settings are not described. The paper is potentially interesting but requires substantial additional detail and experiments before its claims can be accepted.","major_comments":[{"comment":"The reported baseline OA is 93.48, which is 9.65 to 12.91 points higher than the OA of every semi-supervised comparison method (LSST 80.57, ClassHyPer 83.62, ST++ 83.83), and SAMST's OA is 86.44, a 7.04-point drop from the baseline despite higher mIoU and higher per-class F1 for Building, Tree, and Surface. Since Building, Tree, and Surface dominate the Potsdam scene, this pattern is internally inconsistent under a matched evaluation protocol. The authors should explain the experimental setup that produces this pattern, report per-class recall and confusion matrices, and verify that all methods use the same test split, evaluation code, backbone, and training budget. Without this, the reported mIoU gain cannot be attributed to SAMST.","section":"Table I, Section III.C"},{"comment":"The core refiner depends on seven prompt-generation parameters (Bn, Pp, Pn, Tp, Pm, Tn, Nm) and two stitching thresholds (tc, to), but the paper only states that these are 'set through experimental outcomes' and gives neither their values nor the procedure used to choose them. The same omission applies to alpha in Eq. (1) and the class weights wl,c and wp,c in Eqs. (2)-(3). Without these details the method is not reproducible, and there is no way to assess whether the choices were made on a validation set or the test set. Please report the chosen values and the validation criterion used.","section":"Section II.A"},{"comment":"The text says that a checkpoint is saved every 1,000 iterations and that 'the best checkpoint' is used to generate pseudo-labels, but the selection criterion is not defined and no validation split is described. If the test set was used to select among the 80 saved checkpoints, the reported numbers are optimistically biased. Please specify the validation protocol and the metric used for checkpoint selection, and apply the same protocol to the baseline and comparison methods.","section":"Section III.B"},{"comment":"No ablation study is provided, so it is impossible to determine which component of the proposed framework (Threshold Filter Module, Prompt Generation Module, Label Refinement Module, or the composite loss) is responsible for the reported improvement. In addition, the paper claims an iterative framework but reports only a single full iteration. Please provide module-wise ablations and at least one multi-iteration experiment to support the central claim.","section":"Section III.C and Table I"},{"comment":"The implementation details are given only for SAMST; there is no description of how LSST, ST++, and ClassHyPer were configured for comparison. It is therefore unclear whether the comparison methods used the same backbone, the same labeled/unlabeled split, the same input resolution, the same optimizer settings, and the same training iterations. Without matched settings, the relative gains in Table I could be an artifact of weaker baselines. Please provide full training and evaluation details for all compared methods, and ideally repeat experiments over multiple seeds to report variance.","section":"Section III.B"}],"minor_comments":[{"comment":"The text in panel (d) labels the class as 'Imprevious surfaces'; this should be 'Impervious surfaces'.","section":"Fig. 2"},{"comment":"The dataset split is ambiguous: 'we allocated 24 images for training and 14 for validation and testing' is followed by a statement that the testing set includes 2016 patches. Please clarify how the validation set and test set are separated, and whether the 2016 patches come from the 14 images or from a different partition.","section":"Section III.A"},{"comment":"The Label Refinement Module description says holes are removed from each mask to obtain mask m1, but the paper does not define what 'holes' means concretely or how the masks from all connected regions are concatenated. Please provide a precise algorithmic description or pseudo-code.","section":"Section II.A.3"},{"comment":"There is no dedicated related-work section; the introduction mentions consistency regularization and self-training but does not discuss the cited remote sensing semi-supervised methods in enough detail to position the novelty of SAMST. Consider adding a short related-work section.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The central claim is not yet supported because of the unexplained OA pattern in Table I and the undisclosed hyperparameters. I do not see evidence of circular reasoning or fabricated results, and the main issues are addressable with additional documentation and experiments, so a major revision is more appropriate than rejection. The authors should be required to clarify the evaluation protocol, report all threshold values and selection procedures, and provide ablations and matched baselines."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nThe short version: this is a plausible engineering recipe for using SAM to clean up pseudo-labels in remote sensing semi-supervised segmentation, but the main evidence table is self-undermining. The method description is clear enough to implement, and the specific Prompt Generation Module — connected components, prompt boxes, positive/negative points — is a concrete extension of the SAM-as-Guide idea in [19]. The composite loss is standard. On the positive side, the visualizations show the refiner can separate stuck cars, fix boundaries, and mark uncertain regions as ignored; that qualitative story is believable.\n\nThe problem is Table I. The reported baseline reaches OA 93.48, while every semi-supervised comparison sits between 80.57 and 83.83, and SAMST itself drops to 86.44. Meanwhile, per-class F1 for Building, Tree, and Surface all improve, and Clutter improves by 8 points. If those large classes improve, a 7-point OA drop is not credible unless there is a catastrophic drop in an unreported class. The paper does not mention this. The mIoU gain of 5.67 points over the baseline could be real, but the table as presented makes it hard to trust. Add the missing error bars, the absence of ablations for the three modules, the undisclosed values for the seven prompt thresholds plus tc and to, and the unstated criterion for the best checkpoint, and the central claim becomes under-evidenced. The comparison numbers from LSST, ST++, and ClassHyPer are likely taken from prior papers without matched backbones or protocols, which only adds to the confusion.\n\nThere is no code, and the hyperparameters are said to be set through experimental outcomes, which is exactly the kind of hand-tuning that needs to be reported against a validation split. If any threshold or checkpoint choice used the test set, the reported numbers are optimistically biased.\n\nSo the idea deserves further work, but the write-up is not ready. It could be a useful contribution after matched comparisons, ablations, hyperparameter disclosure, and an explanation of the OA discrepancy. As it stands, the evidence table does not hang together.\n\nFor you: if you work on remote sensing semi-supervised segmentation, this is worth a look as a recipe, but I would not cite it yet. I would send it to a serious referee, but with the expectation that a major revision is needed. The method is not nonsensical; the reporting is too weak to verify.\n\nBest,\n[Your name]","headline":"Plausible SAM-based pseudo-label refinement recipe, but Table I is internally inconsistent and the evidence is under-specified.","tokens_in":7527,"tokens_out":7088,"would_cite":false,"duration_ms":76227,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SAMST claims that a frozen SAM can be used to clean noisy pseudo-labels, lifting semi-supervised segmentation on Potsdam to 70.80 mIoU from a 67.00 baseline with only 1/32 of labels.","keywords":["semi-supervised semantic segmentation","remote sensing","pseudo-label refinement","Segment Anything Model","self-training","land cover segmentation","transformer","Potsdam dataset"],"falsifier":"Take the fully trained SAMST pipeline and run the refiner with SAM's masks replaced by the thresholded connected regions themselves (i.e., no SAM re-segmentation), keeping all thresholds identical. If the mIoU gain over the baseline nearly disappears, the gain is attributable to SAM's masks; if the gain persists, the filtering/stitching alone is responsible. Alternatively, measure the pixel accuracy of refined pseudo-labels against the ground truth on the held-out portion of unlabeled data: if the refinement does not raise accuracy relative to raw predictions, the reported gains cannot come from improved label quality.","tokens_in":6413,"feed_emoji":"🛰️","tokens_out":6202,"duration_ms":58440,"temperature":0.7,"pith_summary":"SAMST claims that a frozen Segment Anything Model (SAM) can be used as a post-hoc refiner to clean the noisy pseudo-labels that a weakly supervised segmentation model produces, and that this lets semi-supervised remote sensing segmentation work well with just 1/32 of the training data labeled. On the Potsdam dataset, the method reports mIoU 70.80, OA 86.44, and mF1 81.65 after a single self-training iteration, beating its own supervised baseline by 5.67 mIoU points and outperforming three published semi-supervised methods. The paper argues that the gain comes from improving pseudo-label accuracy rather than from more training, because the refiner removes low-confidence predictions, prompts SAM to redraw class-agnostic object boundaries, and stitches the corrected masks back into the label map with probability-based thresholds. A composite weighted loss that downweights pseudo-labeled samples is introduced to keep label noise from derailing the model.","feed_headline":"SAM refiner adds 5.7 mIoU at 1/32 label ratio","feed_subtitle":"SAMST's SAM-based Pseudo-label Refiner cleans model predictions, beating LSST, ST++, and ClassHyPer on Potsdam with 1/32 labels.","key_machinery":"The load-bearing mechanism is the SAM-based Pseudo-label Refiner, a three-module pipeline: (1) the Threshold Filter Module zeroes out low-confidence predictions into an ignored label (255); (2) the Prompt Generation Module converts the filtered map into connected regions, computes maximum bounding rectangles expanded by Bn pixels, and samples positive/negative points with proximity and probability constraints to prompt SAM; (3) the Label Refinement Module takes SAM's masks, removes holes, and stitches them with the model's probability map using thresholds tc and to, updating or ignoring each pixel. This refiner is what converts SAM's class-agnostic, boundary-sensitive masks into class-aware pseudo-labels, and it is the component whose ablation would separate the paper's contribution from ordinary self-training.","core_discovery":"The paper's central claim is that the SAM-based Pseudo-label Refiner, composed of a Threshold Filter Module, a Prompt Generation Module, and a Label Refinement Module, turns noisy model predictions into substantially more accurate pseudo-labels, and that this refinement is the reason SAMST improves semi-supervised segmentation. The refiner first masks out pixels whose predicted class probability falls below class-specific thresholds, then finds connected regions for each surviving class, builds prompt boxes around them, places positive and negative points inside and outside the regions, and feeds these prompts to SAM to obtain object masks. The Label Refinement Module then assigns a pixel to a class only when the SAM mask's category agrees with the model's high-confidence prediction or when the competing class is very unlikely, otherwise marking the pixel as ignored. The paper reports that after one iteration this pipeline lifts mIoU from 67.00 to 70.80 on Potsdam at a 1/32 label ratio, surpassing LSST, ST++, and ClassHyPer.","pith_inferences":["A controlled ablation that replaces SAM's masks with class-agnostic masks of identical coverage would reveal how much of the mIoU gain is due to SAM's boundary accuracy versus simple spatial smoothing; the paper does not report this ablation.","The 255 ignore-label mechanism may act as a selective curriculum: by injecting a modest number of (or zero) pseudo-labels for difficult regions, the model avoids reinforcing its own errors. This could explain the large clutter-class gain and is testable by measuring training-set class distribution before and after refinement.","The same refiner could be plugged into consistency-regularization methods, not just self-training, since it only touches pseudo-labels; testing SAMST's refiner inside an ST++-style pipeline would isolate whether the gains come from the refiner or the overall two-stage schedule.","Dataset-specific thresholds (Bn, Pp, Pn, Tp, Pm, Tn, Nm, tc, to) suggest the method may transfer less cleanly to other resolutions or class taxonomies; a sensitivity analysis over these parameters would clarify how much tuning is needed."],"forward_implications":["SAM-based pseudo-label refinement yields a 5.67-point mIoU gain over the supervised baseline at a 1/32 label ratio on Potsdam.","A single self-training iteration suffices to surpass the three compared semi-supervised methods (LSST, ST++, ClassHyPer) on the same setup.","Per-class F1 rises for five of six classes, with the largest gain (17.68 points) on the clutter class; low vegetation decreases slightly.","Because the refiner is prompt-based and class-agnostic, the same SAMST framework can be applied to other remote sensing datasets without retraining SAM, as long as the prompt and threshold hyperparameters are re-tuned."],"supporting_citations":[{"why":"Supplies the Segment Anything Model, the class-agnostic promptable segmenter whose masks the Pseudo-label Refiner uses for boundary refinement and error removal.","marker":"[18]"},{"why":"Provides the Potsdam dataset, the benchmark on which the central mIoU 70.80 result is measured at a 1/32 label ratio.","marker":"[20]"},{"why":"ST++ is the strongest self-training baseline the method is compared against and whose reported mIoU (64.89) is surpassed.","marker":"[11]"},{"why":"LSST is the remote sensing self-training baseline defining the previous standard that SAMST outperforms.","marker":"[12]"},{"why":"ClassHyPer is the consistency-perturbation baseline that the method exceeds, anchoring the comparison on the Potsdam setup.","marker":"[2]"},{"why":"Establishes the pseudo-label self-training paradigm that SAMST's two-stage supervised/pseudo-label training builds on.","marker":"[17]"},{"why":"Prior work using SAM as a pseudo-label refiner for referring expression segmentation, the direct methodological precursor the paper adapts to remote sensing.","marker":"[19]"}],"fun_headline_variants":["SAM pseudo-label refiner boosts mIoU by 5.7","SAMST: SAM refines pseudo-labels for remote sensing","1/32 labels: SAM refiner adds 5.7 mIoU","Pseudo-label refinement with SAM adds 5.7 mIoU"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's gain rests on the assumption that SAM's boundary-preserving masks, when prompted by the noisy model's boxes and points and stitched back with the probability thresholds, improve pseudo-labels instead of injecting new errors.","fun_headline_variants_meta":{"raw":{"variants":["SAM pseudo-label refiner boosts mIoU by 5.7","SAMST: SAM refines pseudo-labels for remote sensing","1/32 labels: SAM refiner adds 5.7 mIoU","Pseudo-label refinement with SAM adds 5.7 mIoU"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001078,"raw_usage":{"total_tokens":4513,"prompt_tokens":948,"completion_tokens":3565,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":3486}},"tokens_in":564,"tokens_out":3565,"duration_ms":28220,"temperature":1.0,"reasoning_tokens":3486,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:55:43.913284+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the fully trained SAMST pipeline and run the refiner with SAM's masks replaced by the thresholded connected regions themselves (i.e., no SAM re-segmentation), keeping all thresholds identical. If the mIoU gain over the baseline nearly disappears, the gain is attributable to SAM's masks; if the gain persists, the filtering/stitching alone is responsible. Alternatively, measure the pixel accuracy of refined pseudo-labels against the ground truth on the held-out portion of unlabeled data: if the refinement does not raise accuracy relative to raw predictions, the reported gains cannot come from improved label quality.","supporting_citations":[{"cited_title":"The isprs benchmark on urban object classification and 3d building reconstruction,","cited_arxiv_id":null,"evidence_quote":"Provides the Potsdam dataset, the benchmark on which the central mIoU 70.80 result is measured at a 1/32 label ratio."},{"cited_title":"St++: Make self- training work better for semi-supervised semantic segmenta- tion,","cited_arxiv_id":null,"evidence_quote":"ST++ is the strongest self-training baseline the method is compared against and whose reported mIoU (64.89) is surpassed."},{"cited_title":"Simple and efficient: A semisupervised learning framework for remote sensing image semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"LSST is the remote sensing self-training baseline defining the previous standard that SAMST outperforms."},{"cited_title":"Classhyper: Classmix-based hybrid perturbations for deep semi-supervised semantic segmentation of remote sensing imagery,","cited_arxiv_id":null,"evidence_quote":"ClassHyPer is the consistency-perturbation baseline that the method exceeds, anchoring the comparison on the Potsdam setup."},{"cited_title":"Pseudo-label: The simple and efficient semi- supervised learning method for deep neural networks,","cited_arxiv_id":null,"evidence_quote":"Establishes the pseudo-label self-training paradigm that SAMST's two-stage supervised/pseudo-label training builds on."},{"cited_title":"SAM as the Guide: Mastering Pseudo-Label Refinement in Semi-Supervised Referring Expression Segmentation","cited_arxiv_id":"2406.01451","evidence_quote":"Prior work using SAM as a pseudo-label refiner for referring expression segmentation, the direct methodological precursor the paper adapts to remote sensing."}],"review_version":1}