{"id":"4b4c19dd-592a-44c9-be54-441b2413aaaf","arxiv_id":"2506.10503","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A two-stage framework, PSLG-SAM, uses visual grounding to locate objects and Segment Anything Model to segment them, with a new RRSIS-M dataset, surpassing prior methods when mask supervision is available.","lead":"This paper builds a two-stage system for remote sensing image segmentation from text descriptions: a visual grounding module finds the object, then Segment Anything Model cuts it out. The authors add a new annotated benchmark (RRSIS-M) and report state-of-the-art results on two datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The RRSIS-M SOTA claim is confounded: RRSIS-M is built directly from DIOR-RSVG, the same source used to train LQVG, so the grounding box prompts may encode memorized test annotations.","rationale":"The reader's weakest assumption already identified the unvalidated transfer of LQVG and the possible overlap between RRSIS-M and LQVG's pretraining data. My stress-test sharpens this into a concrete, document-supported concern: RRSIS-M is explicitly constructed from DIOR-RSVG (Section 4.1, Step 1), and LQVG is a DIOR-RSVG-trained grounding model (Section 2.2). Therefore the condition required for the central claim to hold — that the RRSIS-M benchmark measures genuine generalization of the grounding stage — is not established. The paper's own text supplies the evidence for this concern; it is not an external assumption. Because the reader's verdict is already CONDITIONAL, and the required condition is precisely the split-independent evaluation of LQVG, my read does not change the verdict. If the proposed split check shows zero overlap and comparable LQVG accuracy on unseen RRSIS-M images, the central claim would be substantially supported; if overlap is found, the RRSIS-M results should be discounted or re-reported with a disjoint LQVG training set. The RRSIS-D fully supervised result is less directly affected, but the same check should be run for completeness.","tokens_in":12011,"tokens_out":3608,"duration_ms":42618,"concrete_test":"Obtain or reconstruct the exact DIOR-RSVG train split used to train the adopted LQVG checkpoint and compare image filenames or image IDs against the RRSIS-M test set (1,104 images) and the RRSIS-D test set. Report the overlap count and per-category breakdown. If any overlap exists, retrain or re-select LQVG using only DIOR-RSVG images disjoint from the RRSIS-M test set (ideally also disjoint from RRSIS-D test images), then rerun Tables 1 and 2 and recompute oIoU, mIoU, and P@X. If RRSIS-M is a re-split of DIOR-RSVG, this comparison is the decisive check for whether the claimed state-of-the-art results survive without train/test leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim that PSLG-SAM surpasses state-of-the-art methods on RRSIS-M depends on a clean evaluation of the LQVG grounding stage. Section 4.1 states that RRSIS-M was constructed by \"leveraging bounding box prompts and textual descriptions derived from the DIOR-RSVG dataset,\" while Section 2.2 says the framework \"adopts LQVG\" for coordinate prompts, and LQVG is a visual grounding model trained on DIOR-RSVG. The paper never reports the DIOR-RSVG train/test split used for LQVG, nor does it state whether the RRSIS-M test images (and their text-box pairs) were excluded from LQVG's training data. If the RRSIS-M test images appeared during LQVG training, then the bounding box coordinates fed to CFPG and SAM in Tables 2 and 4 are not produced by genuine grounding transfer; they can reflect memorized annotations. This is especially damaging to the weakly supervised setting, where the box from LQVG is effectively the only supervision signal at test time, and to the claim that the second stage can be train-free. The same overlap concern also applies to RRSIS-D if its source imagery overlaps with DIOR-RSVG, which the paper does not discuss. The comparison fairness is further strained because the baseline methods are not given access to LQVG's pretrained grounding knowledge, so even without overlap the comparison conflates the contribution of the proposed framework with the contribution of a separately pretrained grounding model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes PSLG-SAM, a two-stage framework for reference remote sensing image segmentation (RRSIS). In the first stage, a visual grounding model (LQVG) converts an image-text pair into a bounding-box coordinate prompt. In the second stage, the bounding box crops the region of interest, a clustering-based foreground point generator (CFPG) supplies a center point, and the Segment Anything Model (SAM) with a frozen image encoder and prompt encoder produces a mask; a mask boundary iterative optimization (MBO) strategy, inspired by GrabCut, refines the mask edges. The mask decoder can be fine-tuned or kept frozen, yielding fully supervised and weakly supervised variants. The paper also introduces RRSIS-M, a manually annotated benchmark derived from the DIOR-RSVG dataset, and reports experiments on RRSIS-D and RRSIS-M that claim state-of-the-art performance in both settings.","tokens_in":12396,"tokens_out":5473,"duration_ms":64080,"significance":"The two-stage decomposition is conceptually attractive and practically relevant: it leverages strong pretrained components (visual grounding and SAM) and offers a train-free segmentation stage that reduces mask-annotation demand. The reported RRSIS-D fully supervised mIoU improvement (70.61 vs. 64.22 for RMSIN) is substantial, and the RRSIS-M dataset, if carefully constructed, could be a useful contribution to the community. The paper also provides ablation evidence that CFPG and MBO contribute to the final scores, and it promises public code. However, the central claim of state-of-the-art performance is currently conditional on an unverified data-separation assumption: RRSIS-M is built from DIOR-RSVG, and LQVG is a DIOR-RSVG-era visual grounding model, so the evaluation may involve train/test overlap in the grounding stage. Until this is resolved, the significance of the empirical results is not fully established.","major_comments":[{"comment":"The RRSIS-M evaluation has a potential train/test overlap that is not addressed. Section 4.1 states that RRSIS-M was constructed by \"leveraging bounding box prompts and textual descriptions derived from the DIOR-RSVG dataset,\" while Section 2.2 says the framework \"adopts LQVG\" to generate coordinate prompts, and LQVG is a visual grounding model developed on DIOR-RSVG. The paper does not report the DIOR-RSVG split used to train LQVG, nor does it state whether the RRSIS-M test images and their text-box pairs were excluded from LQVG's training data. If they were not excluded, the bounding boxes fed to CFPG and SAM in Tables 2 and 4 are not produced by genuine grounding transfer, and the weakly supervised results—where the LQVG box is the only test-time localization signal—would be invalid as evidence for the method. The same concern applies to RRSIS-D if its source imagery overlaps with DIOR-RSVG, which the paper also does not discuss.","section":"§2.2 and §4.1"},{"comment":"The comparison with state-of-the-art methods is not confound-controlled. PSLG-SAM uses a pretrained SAM and a pretrained LQVG, while the baseline methods (RRN, CMSA, LAVT, RMSIN, etc.) are trained from scratch on the task-specific mask labels. The reported SOTA claim therefore conflates the contribution of the proposed framework with the contribution of additional pretrained grounding knowledge and SAM's large-scale pretraining. A controlled experiment—for example, feeding LQVG's bounding boxes as external prompts to a baseline segmentation model, or fine-tuning all methods with the same pretrained backbone—is needed to attribute the performance gain to the proposed components.","section":"§4.4, Tables 1 and 2"},{"comment":"The weakly supervised claim is only partially supported on RRSIS-D. In Table 1, PSLG-SAM* achieves mIoU 67.41 versus 64.22 for RMSIN, but its oIoU is 75.40 versus 77.53, i.e., 2.1 points lower. The text says the method achieved \"a 3.2% improvement in mIoU and a 2.1% difference in oIoU\" and later acknowledges the lower oIoU, but the abstract and conclusion present the weakly supervised variant as a success without this caveat. The authors should state plainly that the weakly supervised model improves mIoU but not oIoU on RRSIS-D, and justify why mIoU should be treated as the primary metric for the SOTA claim.","section":"§4.4, Table 1"},{"comment":"The CFPG module relies on the assumption that the target foreground object is the connected component with the greatest convexity within the bounding box. This is stated as an observation that \"almost all\" foreground objects have smaller convexity defects than background regions, but no quantitative validation is provided. For concave or complex targets such as an expressway service area or scattered buildings, the largest-convexity selection could pick a background region. A failure analysis or an experiment with oracle bounding boxes would clarify how often this assumption holds and how much of CFPG's reported gain depends on it.","section":"§3.2, Eq. (8)"}],"minor_comments":[{"comment":"There are several typos, including \"aassist\" in the contribution list and \"RRISS\" in Figure 1; these should be corrected.","section":"Abstract and Introduction"},{"comment":"The mask boundary optimization is described as an energy minimization with a data term and a smoothness term, but the smoothness term is never written as an equation and the weighting between the terms is not specified. The authors should provide the full energy function and the parameter values for reproducibility.","section":"§3.3"},{"comment":"It is unclear whether LQVG is trained from scratch or fine-tuned from a released checkpoint, and on which data split. This is directly relevant to the data-leakage concern and should be stated explicitly.","section":"§4.2"},{"comment":"The column header \"Pr@0.5\" appears as \"P r@0.5\" and the metric is sometimes written \"MIoU\" and sometimes \"mIoU\"; the caption should define all abbreviations consistently.","section":"Table 1 and Table 2"},{"comment":"The sentence describing the weakly supervised RRSIS-D result as a \"2.1% difference in oIoU\" is confusing because the difference is a decrease, not an improvement; please rephrase to state the direction explicitly.","section":"§4.4"},{"comment":"The conclusion claims the framework is a \"scalable, efficient solution\" and \"offering promising potential,\" but it does not mention the reliance on LQVG's pretrained knowledge or the possibility that LQVG boxes can be inaccurate for unusual remote sensing targets. A limitation paragraph would improve the paper's balance.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"The main risk is data leakage from LQVG's pretraining into the RRSIS-M test set, since RRSIS-M is constructed from DIOR-RSVG and LQVG is a DIOR-RSVG visual grounding model. I would treat the exclusion of RRSIS-M test images (and RRSIS-D, if applicable) from LQVG's training data as a gate for the revision. If the authors cannot provide a disjoint training split, the headline SOTA claim is not supportable and the paper would need to be restructured around the RRSIS-D results with a clearly stated overlap limitation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the two-stage design (grounding then SAM) is a natural and well-executed idea, and RRSIS-M looks like a genuinely useful dataset. But the headline 'surpasses SOTA' is too strong as it stands: the RRSIS-M numbers are confounded because the grounding model is trained on the same source data, and the fully supervised gains on RRSIS-D need a controlled comparison against baselines that get the same box prompts.\n\nWhat's new and good: the first explicit decomposition of RRSIS into visual grounding plus SAM segmentation. CFPG and MBO are clearly described, and the ablation shows each helps. The train-free second stage with optional decoder fine-tuning is practical. RRSIS-M, with multi-category manual annotations and quality control, is a real contribution even if the evaluation is questionable. The fully supervised result on RRSIS-D (mIoU 70.61 vs 64.22 for RMSIN) is a large gain worth taking seriously.\n\nSoft spots, in order of severity:\n1. The RRSIS-M evaluation is not clean. RRSIS-M is built from DIOR-RSVG, and LQVG is trained on DIOR-RSVG. The paper never says whether the test images and text-box pairs were excluded from LQVG's training. If they weren't, the boxes in Tables 2 and 4 are memorized, not transferred. This is especially damaging for the weakly supervised results, where the box is effectively the only supervision. The authors must state the split and, better, retrain or report results using a grounding model not trained on that distribution.\n2. The comparison is not controlled. Baselines are trained from scratch; PSLG-SAM gets pretrained LQVG and SAM. Some of the gain may just be the pretrained grounding. A fair test would give baselines the same box prompts or train LQVG on a disjoint set. Without that, 'surpasses SOTA' is not established.\n3. Weakly supervised oIoU actually drops on RRSIS-D: 75.40 vs RMSIN's 77.53. The paper acknowledges this in Section 4.4, but the abstract's 'surpasses existing state-of-the-art' is too broad.\n4. Minor: free parameters (k, area threshold, watershed threshold, GMM components) are not sensitivity-analyzed, and the code is promised but not released.\n\nThe math and derivations are coherent. The concerns are about evaluation hygiene, not core validity. A serious referee should engage; the paper needs revision, not rejection. Recommend sending to peer review with requests for split disclosure, controlled comparisons, and code.","headline":"A sensible two-stage framework with a genuinely useful new dataset, but the RRSIS-M SOTA claim is confounded by LQVG's DIOR-RSVG training; needs controlled comparisons before accepting the headline.","tokens_in":12893,"tokens_out":2278,"would_cite":false,"duration_ms":26276,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-stage pipeline that converts a text description into a bounding box and lets the Segment Anything Model refine the mask outperforms end-to-end referring remote sensing segmentation models.","keywords":["reference remote sensing image segmentation","visual grounding","Segment Anything Model","prompt-guided segmentation","weakly supervised segmentation","GrabCut boundary refinement","RRSIS-M dataset","clustering-based foreground point generator"],"falsifier":"Measure LQVG's box localization accuracy (for example, box IoU against the ground-truth mask's bounding box) on the RRSIS-D and RRSIS-M test sets, and rerun PSLG-SAM with ground-truth boxes replacing the predicted ones. If the ground-truth-box version is substantially better, the claimed state-of-the-art result is carried by the grounding stage; if it is not, the segmentation stage is the limiting factor. In addition, checking the overlap between RRSIS-M test images and LQVG's DIOR-RSVG training images would settle whether the new-dataset numbers measure transfer or memorization.","tokens_in":11817,"feed_emoji":"🛰️","tokens_out":8641,"duration_ms":86979,"temperature":0.7,"pith_summary":"Reference remote sensing image segmentation asks a model to produce a segmentation mask for the object a sentence describes. The paper's central claim is that this task does not need an end-to-end multimodal segmentation network: it can be split into a coarse localization stage, where a visual grounding model converts the text into bounding-box coordinates, and a fine segmentation stage, where the Segment Anything Model (SAM) turns the box into a precise mask. This second stage can be train-free and needs no mask-level labels, so the framework's weakly supervised variant only requires bounding boxes. On the standard RRSIS-D benchmark the fully supervised version lifts mIoU from the previous state of the art's 64.22 to 70.61, and the same two-stage design also reports gains on the paper's new manually annotated RRSIS-M dataset. A sympathetic reader would take the contribution to be a demonstration that decomposing the task by semantic localization is both more accurate and much cheaper in annotation effort than single-stage fusion.","feed_headline":"Grounding-first SAM pipeline tops referring segmentation benchmarks","feed_subtitle":"A two-stage model turns text into a box, then SAM refines the mask, beating end-to-end networks without dense mask labels.","key_machinery":"The machinery is a prompt chain from text to mask. Stage one is LQVG, a DETR-style visual grounding model that fuses text and image features and outputs a bounding box around the described object. Stage two is SAM with a frozen image encoder, whose prompt encoder receives that box plus a foreground point produced by the clustering-based foreground point generator (CFPG): CFPG crops the box, runs KMeans++ to split pixels into two clusters, applies watershed segmentation to separate connected regions, and selects the region with the highest convexity as the target, then takes its geometric center as the point prompt. The SAM mask is then refined by a mask boundary iterative optimization strategy adapted from GrabCut, which builds foreground and background Gaussian mixture models and minimizes a graph-cut energy. The argument is that this chain lets SAM's boundary competence carry the fine segmentation while the grounding stage supplies the semantic localization that end-to-end networks struggle to learn from scarce remote-sensing annotations.","core_discovery":"The central discovery claimed is that referring remote sensing segmentation can be reframed as localization-plus-segmentation: a remote-sensing visual grounding model (LQVG) provides prompt coordinates, and a frozen SAM, guided by a clustering-based foreground point generator and a mask boundary iterative optimization strategy, produces the final mask. The framework is reported to surpass existing state-of-the-art models on both RRSIS-D and the new RRSIS-M dataset, with the weakly supervised variant (trained on bounding boxes only) already beating the prior fully supervised RMSIN baseline in mIoU on RRSIS-D (67.41 versus 64.22). Fine-tuning only the lightweight SAM mask decoder on 10% of the mask data recovers most of the fully supervised gain (70.13 versus 70.67 mIoU in the paper's data-size experiment). A t-SNE analysis of encoder outputs is offered as evidence that SAM's frozen features separate remote-sensing categories more cleanly than the multi-scale features of the previous best method.","pith_inferences":["An experiment the paper does not run would apportion credit: replace LQVG boxes with ground-truth boxes of matched tightness; a large mIoU jump would show the reported gains are mainly the grounding stage's contribution rather than SAM's or the prompt modules'.","The paper's RRSIS-M test set is built from DIOR-RSVG imagery and descriptions, while LQVG is trained on DIOR-RSVG; checking whether test images overlap the grounding training split would clarify whether the RRSIS-M results reflect generalization or near-duplicate transfer.","The convexity heuristic inside CFPG assumes the target is the most convex connected region in the box, so thin, elongated, or partially occluded objects are a plausible failure mode worth a targeted stress test.","If the two-stage recipe transfers beyond the specific models used here (for example, different grounding backbones or different SAM checkpoints), it would suggest that referring segmentation in general domains can be decomposed into localization and boundary refinement rather than learned end-to-end."],"forward_implications":["Referring remote sensing segmentation can be performed with bounding-box-only supervision; the weakly supervised variant already exceeds the prior fully supervised baseline in mIoU on RRSIS-D (67.41 versus 64.22).","The segmentation stage can be kept frozen, so adapting the framework to a new region or sensor costs only the grounding model, not a mask-training run.","Fine-tuning only the lightweight SAM mask decoder on 10% of the mask data recovers most of the gain of full supervision (70.13 versus 70.67 mIoU at 100% in the paper's data-size experiment), so large mask-annotation campaigns have diminishing returns once localization is solved.","Because the stages are independent, improving the visual grounding model or the prompt-generation modules should directly improve final masks without re-designing the segmentation pipeline."],"supporting_citations":[{"why":"Supplies the RRSIS-D benchmark, the RMSIN state-of-the-art baseline, and the prior numbers the paper must beat.","marker":"[Liu et al.(2024)]"},{"why":"Introduced the RRSIS task and its first dataset, defining the problem setting the paper addresses.","marker":"[Yuan et al.(2024)]"},{"why":"Provides DIOR-RSVG, the remote-sensing grounding dataset whose imagery and text descriptions are used to build RRSIS-M and which grounds the localization stage.","marker":"[Zhan et al.(2023)]"},{"why":"The DETR architecture that LQVG is described as being inspired by, anchoring the detection-style coordinate output design.","marker":"[Carion et al.(2020)]"},{"why":"The GrabCut algorithm that the mask boundary iterative optimization strategy directly adapts.","marker":"[Rother et al.(2004)]"},{"why":"The KMeans++ seeding used by the clustering-based foreground point generator to split the cropped box into foreground and background clusters.","marker":"[Arthur and Vassilvitskii(2007)]"},{"why":"The watershed segmentation used by the foreground point generator to separate connected regions before convexity selection.","marker":"[Meyer(1992)]"},{"why":"LAVT, a strong referring-segmentation baseline that the framework is compared against on both datasets.","marker":"[Yang et al.(2022)]"}],"fun_headline_variants":["Two-stage PSLG-SAM tops referring remote sensing segmentation","Grounding-then-SAM pipeline beats end-to-end for remote sensing","SAM with localization prompts beats SOTA in remote sensing segmentation","Coarse localization guides train-free SAM for precise remote segmentation","Weakly supervised PSLG-SAM outperforms fully supervised baseline"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework assumes the pretrained LQVG visual grounding model already knows how to find the text-described object and frame it with a usable box in remote sensing images; the paper never measures LQVG's grounding accuracy on RRSIS-D or RRSIS-M, and if those boxes drift, the clustering point generator and SAM cannot recover.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage PSLG-SAM tops referring remote sensing segmentation","Grounding-then-SAM pipeline beats end-to-end for remote sensing","SAM with localization prompts beats SOTA in remote sensing segmentation","Coarse localization guides train-free SAM for precise remote segmentation","Weakly supervised PSLG-SAM outperforms fully supervised baseline"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00081,"raw_usage":{"total_tokens":3578,"prompt_tokens":996,"completion_tokens":2582,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":2505}},"tokens_in":612,"tokens_out":2582,"duration_ms":20467,"temperature":1.0,"reasoning_tokens":2505,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:23:55.739564+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure LQVG's box localization accuracy (for example, box IoU against the ground-truth mask's bounding box) on the RRSIS-D and RRSIS-M test sets, and rerun PSLG-SAM with ground-truth boxes replacing the predicted ones. If the ground-truth-box version is substantially better, the claimed state-of-the-art result is carried by the grounding stage; if it is not, the segmentation stage is the limiting factor. In addition, checking the overlap between RRSIS-M test images and LQVG's DIOR-RSVG training images would settle whether the new-dataset numbers measure transfer or memorization.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The DETR architecture that LQVG is described as being inspired by, anchoring the detection-style coordinate output design."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The KMeans++ seeding used by the clustering-based foreground point generator to split the cropped box into foreground and background clusters."}],"review_version":1}