{"id":"9f53dce2-d2e8-4c4a-97a0-1434d6f747b8","arxiv_id":"2607.16671","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"A foundation-model-assisted active learning system using UPN+SAM2 reference boxes and DINOv2 object features improves low-budget annotation efficiency for remote sensing object detection.","lead":"This paper combines foundation models with active learning to pick which satellite images should be labeled for object detection, and during early training it replaces noisy detector boxes with more stable reference boxes. Experiments on four remote sensing datasets report better detection accuracy at low annotation budgets than several active learning baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SA-source localization reference is unvalidated on dense small/rotated objects; FDSU and DSBS inherit its errors.","rationale":"The reader's weakest assumption named both the SA-source stability and the unmeasured annotation burden. I focus on the SA-source because it is the foundation of the two most novel components (FDSU and DSBS) and directly affects the measured detection performance, not just the secondary cost claim. The paper provides some supporting evidence (Table II) showing UPN+SAM2 improves recall, but that evidence is incomplete: it reports recall at selected thresholds and does not quantify how often SA boxes are worse than detector boxes for the same object. A single well-designed measurement of SA vs OD localization quality on matched objects could settle whether the dual-source mechanism is actually providing a better geometric prior. This concern is not an external-consensus disagreement; it is an internal gap between the assumption and the evidence. The reader already flagged this and issued a CONDITIONAL acceptance; my analysis supports that verdict rather than moving to rejection, because the proposed framework is coherent and the empirical results, while lacking error bars and code, are plausible. I would not change the verdict without the additional check.","tokens_in":13446,"tokens_out":4910,"duration_ms":51457,"concrete_test":"On a held-out validation subset of DOTAv2 and HRSC2016, compute per-object IoU with ground truth for every SA box and the detector's matched OD box (using the same Hungarian matching as Eq. 1). Report the mean IoU and the fraction of matched pairs where the SA box has lower IoU than the OD box. If this fraction is substantial (e.g., >10%) or the mean SA IoU is not clearly higher, the DSBS switching and FDSU localization agreement are not grounded in a better reference, and the cold-start gains from the dual-source components would need a different explanation. Additionally, rerun the low-budget AL on DOTAv2 with SA boxes randomly jittered by ±5 pixels; if mAP drops by more than 2 points, the method is overly sensitive to SA-source noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of cold-start efficiency rests on the assumption that the SA-source (UPN proposals refined by SAM2) provides a stable, detector-independent localization reference. This assumption is load-bearing because both FDSU (Eq. 1, Hungarian matching) and DSBS (Eq. 8, box replacement) depend on the geometric quality of SA boxes. On remote sensing scenes with dense small objects or rotated instances, Table II shows that UPN+SAM2 recall at IoU 0.7 is only 58.8% on DOTAv2 (Top-200), leaving a large fraction of objects without reliable SA boxes. The paper reports recall but not precision or localization accuracy relative to the detector's own boxes. The failure mode is concrete: if an OD box and a false-positive SA box overlap, they produce high localization agreement, leading to low uncertainty in Eq. (2); the sample is then considered uninformative and may not be selected. In DSBS, a thresholded switch (Eq. 8) can replace a reasonably accurate detector box with a worse SA box, adding noise instead of removing it. The paper's Table II does not compare SA box quality to OD box quality on matched objects, and no sensitivity analysis of SA-source degradation is provided. Without such validation, the observed gains cannot be confidently attributed to the dual-source mechanism; they could partly arise from the diversity sampling or from dataset-specific quirks rather than from the SA-source being a better geometric prior.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a foundation-model-assisted active learning and semi-automatic annotation framework for remote sensing object detection. It introduces a dual-source mechanism: an offline-cached 'SA-source' built by UPN proposals refined with SAM2 masks, and an online 'OD-source' from the current detector. Three components are added: FDSU (uncertainty from Hungarian matching between SA and OD boxes plus classification confidence), OCDS (DINOv2 mask-pooled object embeddings with fragmentation suppression), and DSBS (switching noisy detector boxes to SA boxes in cold-start rounds). Experiments on DIOR, HRSC2016, DOTAv2, and FAIR1M compare against Entropy, CoreSet, PPAL, DivProto, and AL4FM, with ablations on DIOR and a foundation-pipeline quality analysis in Table II.","tokens_in":13844,"tokens_out":6899,"duration_ms":71801,"significance":"The core idea—using a reusable, detector-independent localization source to stabilize cold-start active learning—is timely and practically motivated, and the one-time offline caching of SA-source proposals is a sensible design choice. If the reported gains are reproducible, the framework would be a useful step toward reducing annotation cost in remote sensing object detection. The paper also proposes a concrete mechanism (DSBS) to reduce box-editing burden, which is rarely evaluated in active learning papers. However, the current experimental evidence is not yet sufficient: the main comparison lacks error bars, the only tabulated ablation contains a non-monotonic baseline curve, the DSBS section mixes up Recall@100 and mAP, the SA-source is validated only by recall, and no direct measurement of manual annotation burden is provided. These gaps are load-bearing for the central claims of superior sample efficiency and reduced annotation cost.","major_comments":[{"comment":"Fig. 2, the central comparison across the four datasets, is shown without error bars, confidence bands, or the number of seeds. Active-learning pipelines are stochastic in initial labeled selection, detector training, and candidate sampling, so single-run differences of 1–3 mAP ('superior or comparable') are not interpretable. Table I, the only tabulated ablation, contains a red flag: the Entropy row gives mAP 0.5±0.3 at 10 labels and 0.2±0.1 at 50 labels, a non-monotonic drop that suggests a bug, unstable training, or very high variance. Please provide multi-seed statistics for all methods and investigate this baseline.","section":"IV-B / Table I"},{"comment":"The text says 'We evaluate using Recall@100 with a stricter IoU threshold of 75%' but then reports 'mAP' numbers (57%/61%, etc.). Recall@100 and mAP are different metrics; the text must specify which one is plotted in Fig. 3. In addition, the protocol is not described: are the 20/30 annotated samples selected by the AL loop, or is this a separate training run on a fixed small set? Without this information, the claim that DSBS 'substantially improves cold-start performance' cannot be verified.","section":"IV-B, DSBS analysis"},{"comment":"The SA-source is characterized only by proposal recall and an undefined 'AvgΔIoU' column. No precision or object-level localization error relative to ground truth is reported, and there is no comparison of SA-box quality with detector-box quality on matched objects. This matters because FDSU's uncertainty signal (Eq. 2) and DSBS's replacement rule (Eq. 8) both assume SA boxes are a stable geometric reference. On DOTAv2, UPN+SAM2 recall at IoU 0.7 is only 58.8% (Top-200), leaving many objects without an SA box; how these unmatched objects are handled in DSBS is not analyzed. Please define AvgΔIoU, report SA precision/localization accuracy, and add a sensitivity analysis under SA-source degradation.","section":"III-A / III-D / Table II"},{"comment":"The paper motivates DSBS by reducing the manual burden of box refinement, but no direct measurement of annotator effort appears anywhere. mAP or Recall gains are at best indirect proxies. Please add a user study or a quantitative proxy for editing cost—e.g., the fraction of displayed boxes that require correction, the average IoU of displayed boxes with the annotator-accepted box, or the number of mouse/keyboard edits per image—and report it under low-budget conditions.","section":"III-D / IV-B"},{"comment":"In OCDS, the novelty of a candidate object is computed against prototypes of the class predicted by the detector (ĉ_j). In the cold-start regime the detector's classification is itself noisy; if ĉ_j is wrong, the object is measured against the wrong class prototypes and the coverage score is biased. Please analyze sensitivity to label noise or define a class-agnostic novelty fallback.","section":"III-C"}],"minor_comments":[{"comment":"The module is first called 'Dual-Source Box Switching (DBS)' in the Introduction and later 'DSBS' throughout; please use one abbreviation consistently.","section":"I"},{"comment":"The 'AvgΔIoU' column is not defined. Add a formula or a description in the caption so the reader can interpret the values.","section":"Table II"},{"comment":"Reference [9] is cited as UPN, but the listed reference is 'ChatRex: Taming multimodal LLM...' The actual UPN paper appears to be missing; please correct the citation.","section":"References"},{"comment":"The figures have no visible axis labels, legends, or error bars in the provided text. Add these, and consider showing per-seed points for the main comparison.","section":"Figures 2-3"},{"comment":"There are minor typos, e.g., 'and and' in the Introduction, and the expectation notation E[·] in Eq. (10) is introduced without defining the distribution over images and matchings.","section":"Full text"},{"comment":"No code or detailed hyperparameter table is provided. Given the large number of hyperparameters (θ, Ku, Kod, K, Kd, p, λ, τsw, τa, τb, ρmin, ρmax, τupn, τcls, τloc), a supplementary configuration table would greatly aid reproducibility.","section":"Method / Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"Please check the citation of UPN (reference [9] points to ChatRex) and the large number of arXiv 2026 references, some of which appear tangential. I also recommend requesting code, seeds, and error bars as part of the revision. The manuscript is within the journal's scope, but the experimental evidence needs strengthening before the central claims can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a serious engineering paper attacking a real bottleneck: remote sensing annotation cost. The core idea is an offline cache of UPN+SAM2 proposals used as a localization reference (SA-source), matched against the detector's own predictions (OD-source) via Hungarian matching. The uncertainty score fuses localization agreement with classification confidence, the diversity stage uses DINOv2 features masked by SAM2 with an explicit fragmentation penalty, and DSBS swaps noisy detector boxes for SAM2-refined boxes in cold start. That combination is not in the cited baselines, and the framework is coherent and well-motivated.\n\nThe paper does a lot right. It evaluates on four datasets, ablate both stages, and includes a foundation-pipeline quality table. The low-budget gains are plausible, and the fragmentation suppression numbers (34–38% reduction) are concrete.\n\nThe soft spots are mostly about verification, not the idea. No code is released, Fig. 2 has no error bars, and seed counts are missing. More concerning, Table I's Entropy row is non-monotonic (mAP 0.5 at 10 labels, 0.2 at 50) — likely a typo, but it needs fixing. The DSBS manual-burden claim is never directly measured; Recall@100 at IoU 0.75 is a reasonable proxy, but the text confusingly calls it mAP (57%/61% mAP with only 20 samples). Also, the UPN citation [9] points to ChatRex, not UPN — that should be corrected.\n\nThe stress-test concern about SA-source quality is legitimate. Table II shows UPN+SAM2 recall at IoU 0.7 on DOTAv2 is only 58.8% (Top-200). If the SA boxes are wrong, the Hungarian agreement in FDSU and the box replacement in DSBS inherit that noise. The switching threshold tau_sw gives some protection, but the paper does not report SA-box precision or compare SA vs OD localization on matched objects, and there is no sensitivity analysis. This does not sink the paper, but it should be addressed before I'd trust the magnitude of the gains.\n\nVerdict: this deserves a serious referee. It is a well-thought-out framework with real practical potential, but the authors need to release code, fix the citation and the typo, add error bars, and validate the SA-source on dense/rotated scenes. I would not desk-reject.","headline":"Serious engineering with a load-bearing but under-validated SA-source; referee it, but require code, a citation fix, and a sensitivity analysis.","tokens_in":14326,"tokens_out":3659,"would_cite":false,"duration_ms":37768,"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":"The paper claims that a foundation-model-collaborative active learning loop, built on a dual-source mechanism (cached UPN+SAM2 reference boxes plus current detector predictions), achieves superior or comparable detection performance under m","keywords":["active learning","object detection","remote sensing","foundation models","SAM2","UPN","semi-automatic annotation","sample efficiency"],"falsifier":"Measure per-image annotator time or mouse-edit counts on DOTAv2 with DSBS enabled versus disabled; if no reduction in correction effort appears despite improved Recall@75, the annotation-cost claim collapses. A second check: jitter the cached SA boxes randomly and rerun FDSU+OCDS; if low-budget mAP barely changes, the SA-source is not the operative mechanism.","tokens_in":13321,"feed_emoji":"🛰️","tokens_out":6689,"duration_ms":57228,"temperature":0.7,"pith_summary":"This paper argues that the main obstacle to efficient remote sensing object detection annotation is not classification uncertainty but cold-start localization noise: early detectors classify reasonably well while their bounding boxes are still unstable. To fix it, the authors build a dual-source framework: a cached, detector-independent reference pool of boxes and masks produced by UPN followed by SAM2, plus the current detector's own predictions. On top of this they propose FDSU, which fuses localization agreement between the two sources with classification confidence; OCDS, which does mask-guided diversity sampling in DINOv2 feature space while penalizing fragmented proposals; and DSBS, which swaps noisy detector boxes for the reference boxes during early annotation. The paper claims that on DIOR, HRSC2016, DOTAv2, and FAIR1M this combination matches or beats prior active learning methods under most annotation budgets, with the clearest gains in the low-budget cold-start regime. If right, it means fewer images need to be labeled and less manual box correction is needed to reach usable detection performance.","feed_headline":"Dual-source uncertainty sampling beats low-budget detection baselines","feed_subtitle":"A cached UPN+SAM2 reference plus detector predictions rivals or beats prior active learning on four remote sensing datasets.","key_machinery":"The load-bearing mechanism is the SA/OD dual-source alignment. The SA-source is a one-time cached pool of category-agnostic boxes from UPN, refined through SAM2 masks into tight boxes and binary masks; the OD-source is the current-round detector's boxes and class confidences. Hungarian matching with an IoU-based cost (Eq. 1) pairs the two sets; FDSU fuses the matching agreement with classification confidence through a harmonic mean (Eq. 2) to score object-level uncertainty, aggregated by top-K per image. OCDS then builds mask-guided DINOv2 object embeddings, scores novelty against class prototypes, and subtracts a fragmentation penalty; DSBS replaces low-quality detector boxes by matched SA","core_discovery":"The central claim is that a foundation-model-collaborative active learning loop, in which a category-agnostic UPN+SAM2 pipeline serves as a stable spatial reference while the detector supplies category and confidence signals, yields better sample selection and cheaper cold-start annotation than relying on the detector's own predictions alone. The paper asserts that FDSU+OCDS is superior or comparable to existing active learning methods under most annotation budgets across four remote sensing benchmarks, and that the DSBS module substantially improves early-stage Recall@75 at 20-30 labeled images (roughly 57%/61% mAP with DSBS vs 22%/45% without). The authors present this as evidence that loc","pith_inferences":["The paper never directly measures annotator time or correction effort; the claim that DSBS 'reduces manual burden' is inferred from Recall/mAP gains, so a human-in-the-loop timing study is the direct test.","The same dual-source recipe should transfer to other dense-object domains with a category-agnostic proposal model plus a mask generator, such as medical imaging or aerial video, where detector-only confidence is equally unreliable in early rounds.","One robustness gap the authors leave implicit: the SA-source is trusted equally everywhere, so a per-region or per-object trust weighting could prevent false UPN+SAM2 proposals from contaminating FDSU and DSBS on hard scenes.","DSBS scheduling via global matching quality (Eq. 10) could be refined to per-image or per-class switching thresholds, a direction the paper flags only as future work."],"forward_implications":["Low-budget active learning on remote sensing object detection can be made much more sample-efficient; the strongest gains are reported at 10-50 labeled images.","A one-time offline UPN+SAM2 pass can be cached and reused in every active learning round, so the foundation-model inference cost is amortized.","DSBS provides a concrete semi-automatic annotation protocol: geometry from the reference source, labels from the detector, with unmatched boxes handled explicitly.","The framework transfers across diverse remote sensing benchmarks, including dense small objects (DOTAv2) and fine-grained categories (FAIR1M), suggesting the mechanism is not dataset-specific.","Ablations indicate that FDSU and OCDS contribute complementary gains; pairing either with a weak diversity strategy (CoreSet, DivProto) leaves much of the benefit unrealized."],"fun_headline_variants":["Dual-source uncertainty boosts active learning cold-start","Foundation model plus detector beats low-budget detection baselines","Cached UPN+SAM2 reference improves detection annotation efficiency","Active learning with reference source cuts annotation cost","Reference-aided sampling solves cold-start detection annotation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that UPN+SAM2 boxes are a stable, detector-independent localization reference; if they are false positives or poorly aligned on dense, rotated small objects, both FDSU's matching and DSBS's box replacement inject the localization noise they are meant to remove.","fun_headline_variants_meta":{"raw":{"variants":["Dual-source uncertainty boosts active learning cold-start","Foundation model plus detector beats low-budget detection baselines","Cached UPN+SAM2 reference improves detection annotation efficiency","Active learning with reference source cuts annotation cost","Reference-aided sampling solves cold-start detection annotation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000573,"raw_usage":{"total_tokens":2557,"prompt_tokens":768,"completion_tokens":1789,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":1714}},"tokens_in":512,"tokens_out":1789,"duration_ms":12149,"temperature":1.0,"reasoning_tokens":1714,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T20:16:20.141636+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure per-image annotator time or mouse-edit counts on DOTAv2 with DSBS enabled versus disabled; if no reduction in correction effort appears despite improved Recall@75, the annotation-cost claim collapses. A second check: jitter the cached SA boxes randomly and rerun FDSU+OCDS; if low-budget mAP barely changes, the SA-source is not the operative mechanism.","supporting_citations":[],"review_version":1}