{"id":"7c849136-c392-4cd1-aa46-74c4d4cab252","arxiv_id":"2501.13725","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"YOCOv2 adds perceptual-consistency regularization and strong-feature filtering to one-stage YOLO detectors for unsupervised domain adaptation of space terrain detection, claiming up to 31% mAP gains.","lead":"This paper adds two tweaks to a domain-adaptation method for teaching object detectors to recognize craters, dunes, and boulders on Mars, the Moon, and asteroids without labels from the target world. The authors report large accuracy gains on simulated data and show the models can run on spacecraft-like hardware, but the headline 31% improvement is based on one favorable comparison and the code and data are not yet public.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed '31% over YOCOv1' is contradicted by Table III: PC SFF (53.0) is below YOCOv1 (56.1) on the only cited cell; the 31% gap is relative to ViSGA, not YOCOv1.","rationale":"Reader's verdict is CONDITIONAL, and my concern does not move it: the paper still needs major recalibration of its headline claim and external baselines. The load-bearing issue is the central numeric claim: '31% compared with YOCOv1' is not supported by the cited cell and is contradicted by the same table. This is an internal inconsistency, not a disagreement with external consensus, so it directly affects the paper's main contribution. I give credit for the hardware benchmarks (Table I) and for the six-dataset evaluation, and the qualitative results are suggestive, but the core SOTA claim requires correction. The Moon ground-truth incompleteness is honestly flagged by the authors and further undermines the real-world part of the evaluation, but the 31% arithmetic is the more decisive problem. Therefore the verdict remains CONDITIONAL rather than REJECT: the method may still be useful after the claims are fixed and artifacts are provided.","tokens_in":15039,"tokens_out":5425,"duration_ms":43099,"concrete_test":"Recompute from Table III for YOLO 8-S on Sim Asteroid: PC SFF = 53.0, ViSGA = 22.1, YOCOv1 = 56.1. Compare (53.0 - 22.1) = 30.9 against '31%', and note that PC SFF < YOCOv1. If the table is accurate, the abstract's '31% compared with YOCOv1' is false. Once code/data are released, re-train the three methods with identical seeds and splits to confirm these entries and to check whether any other cell yields a 31% advantage over YOCOv1; none is visible in Tables II-IV.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Abstract and conclusion state 'improvements upwards of 31% compared with YOCOv1 and terrestrial state-of-the-art.' The only concrete evidence for this (Section IV-B4, Sim Asteroid, YOLO 8-S, Instance Contrastive) reports PC SFF at 53.0 mAP and ViSGA at 22.1, calling the difference 'nearly 31%'. This is an absolute-point gap of 30.9, a 140% relative gain over ViSGA, not '31% compared with YOCOv1'. Table III shows YOCOv1 at 56.1 mAP for the same architecture and dataset, so the proposed PC SFF is 3.1 mAP below YOCOv1 on this cell. Thus, if '31%' refers to ViSGA, the phrasing is misleading; if it refers to YOCOv1, the paper's own table refutes it. Additionally, 'terrestrial state-of-the-art' is never operationalized: no external DA detection method (e.g., domain-adaptive Faster R-CNN variants) is evaluated, so the SOTA claim has no comparison beyond the authors' own ablations. The paper does admit incomplete Moon labels (Section IV-B5), which weakens the Moon-based support, but the decisive contradiction is the 31% headline itself.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes YOCOv2, an unsupervised domain adaptation (UDA) extension for one-stage YOLO detectors aimed at detecting terrain on Mars, the Moon, and asteroids. It generalizes the prior YOCOv1 architecture by adding a Perceptual Consistency (PC) loss that aligns multi-scale instance features and by introducing Strong Feature Filtering (SFF) and PTAP-based feature selection to make visual-similarity clustering more robust. The paper evaluates 11 loss configurations across 8 YOLO variants on six datasets (Sim Mars, Sim Asteroid, Moon), reports mAP@0.5, performs hardware latency benchmarks on a Zynq/Edge TPU platform, and shows qualitative results on HiRISE, OSIRIS-REx, and lunar imagery. The central claim, repeated in the abstract and conclusion, is that the method improves UDA accuracy by more than 31% compared with YOCOv1 and terrestrial state-of-the-art.","tokens_in":15334,"tokens_out":9186,"duration_ms":77211,"significance":"The proposed pipeline is a reasonable extension of visual similarity-based alignment to one-stage YOLO detectors: the PC term is a natural multi-scale feature-matching regularizer, and the SFF/PTAP mechanisms target a real difficulty in celestial imagery, namely ambiguous feature clusters in texture-poor scenes. If the quantitative claims survive proper evaluation, the method would be a useful step toward onboard UDA for terrain detection. The hardware benchmarking on a Zynq/Edge TPU platform is concrete and relevant to the stated real-time deployment goal, and the breadth of the ablation across three environments and eight YOLO variants is a clear strength. The paper also promises open-source code and datasets. However, the headline result as stated is not supported by the evidence in the paper, and the evaluation lacks multiple seeds and external baselines, so the significance currently rests on a much narrower, and partly contradicted, comparison.","major_comments":[{"comment":"The abstract and Section V claim \"improvements upwards of 31% compared with YOCOv1 and terrestrial state-of-the-art.\" The only concrete evidence cited in Section IV-B4 is the Sim Asteroid / YOLO 8-S / Instance Contrastive cell, where ViSGA scores 22.1 and PC SFF scores 53.0. That is a 30.9 mAP-point gain over ViSGA, not over YOCOv1; the same Table III row lists YOCOv1 at 56.1 mAP, which is 3.1 points above PC SFF. The headline as written is therefore contradicted by the paper's own table. Please re-report this result as an absolute-point improvement over the ViSGA baseline in that cell, and either provide a head-to-head comparison with YOCOv1 or revise the abstract and conclusion to match the actual evidence.","section":"Abstract; Section IV-B4; Table III"},{"comment":"No standard UDA object detector from the terrestrial literature is evaluated. The comparisons in Tables II-IV are restricted to Source Only, Target Only, and internal ablations (ViSGA, PC Only, PC SFF, YOCOv1, PC K-Means, PTAP). Consequently, the \"terrestrial state-of-the-art\" portion of the headline claim is not operationalized and is unsupported. Please add at least one or two established UDA object detection baselines (e.g., DA-Faster R-CNN variants, SWDA, AT, or a one-stage UDA detector) on the same datasets, or restrict the claim to \"compared with the evaluated baselines.\"","section":"Section IV-B; Tables II-IV"},{"comment":"All quantitative results are single-run mAP values with no seeds, error bars, or significance tests. The headline 30.9-point cell in Table III is one configuration among many, and many adjacent entries differ by only 1-3 mAP points, so it is impossible to assess whether the advantages of PC SFF are reliable. Please run at least three seeds per configuration, report mean +/- standard deviation (or confidence intervals), and state whether differences are statistically significant.","section":"Tables II-IV; Section IV-B2"},{"comment":"The Moon evaluation is explicitly acknowledged to be compromised by incomplete ground truth: Section IV-B5 states that the Robbins database does not label every crater, and Figure 5 shows a baseline prediction set that visually exceeds the ground truth. This systematically depresses all Moon mAP values and makes cross-method rankings in Table IV unreliable, yet Moon is one of the three environments used to support the general claim. Please either restrict the Moon evaluation to frames with verified/complete labels, use a curated subset with complete annotation, or relegate Moon results to qualitative/auxiliary evidence and remove them from the quantitative support of the central claim.","section":"Section IV-B5; Figure 5; Table IV"},{"comment":"The proposed modules rely on several hand-set hyperparameters: the agglomerative clustering merging threshold (0.1), the Top-K / top-50% selection threshold for SFF and PTAP, and K=2 for K-Means. No sensitivity analysis is provided. Because SFF and PC are the paper's main contributions, the robustness of the reported gains to these choices should be demonstrated, or at least the values should be justified from validation data rather than presented as fixed constants.","section":"Section III-C; Section IV-B2"}],"minor_comments":[{"comment":"The word \"Adverarial\" is misspelled twice in the qualitative-experiment descriptions; please correct it to \"Adversarial.\"","section":"Section IV-C"},{"comment":"The CPU latency column is nearly constant (22.2-23.1 ms) across models with very different parameter and FLOP counts; please clarify what hardware/software path this column measures and why it does not scale with model size.","section":"Table I"},{"comment":"The sentence \"PC SFF technique outperforms the traditional ViSGA approach in instance clustering methods, surpassing it in four of eight cases\" is ambiguous: in Table III, PC SFF beats ViSGA in 11 of 16 cells if both adversarial and contrastive instance settings are counted, but in only 5 of 8 cells in the contrastive-only setting. Please state exactly which comparison set is being described.","section":"Section IV-B4; Table III"},{"comment":"The description of hierarchical clustering as \"dynamically assigns cluster centers\" is imprecise, since agglomerative clustering with a fixed merging threshold still depends on the chosen distance threshold; please define the clustering procedure and threshold selection precisely.","section":"Section III-C2"},{"comment":"The PTAP selection in Equation (8) is described in words but not fully tied to the Top-K channel selection used for SFF; please provide a precise definition of the set K and how the top-k pixel-wise selection is performed.","section":"Section III-C3"}],"recommendation":"major_revision","confidential_remarks":"I would ask the editor to treat the abstract and conclusion claims as a blocking issue: the paper's own Table III shows PC SFF below YOCOv1 on the only cell cited for the \"31%\" result. The manuscript appears to be an incremental extension of the authors' prior YOCOv1 work, though the hardware benchmark and dataset release are potentially useful. I would also request that the authors clarify the novelty relative to arXiv:2303.04891 and provide a clear statement of which comparisons support each claim in the abstract."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The gist: this is a solid engineering paper with a misleading headline. The new modules—Perceptual Consistency and Strong Feature Filtering—are modest adaptations of existing ideas (multi-scale feature matching, channel attention top-K selection), but combining them in a one-stage YOLO UDA pipeline for celestial terrain is legitimate and the six-dataset evaluation plus the Zynq/Edge TPU hardware benchmark are genuinely useful. The authors are also honest about the Moon ground-truth incompleteness, which is more than many papers do.\n\nThe problem is the central claim. The abstract and conclusion promise \"improvements upwards of 31% compared with YOCOv1 and terrestrial state-of-the-art.\" The only concrete evidence is a 30.9 mAP gap between PC SFF (53.0) and ViSGA (22.1) on Sim Asteroid with YOLO 8-S. But in that same table row, YOCOv1 itself scores 56.1, so PC SFF is actually 3.1 points below YOCOv1. The 31% figure is real only as a gain over ViSGA, not over YOCOv1. And \"terrestrial state-of-the-art\" is never benchmarked—no external domain-adaptive detector appears anywhere in the tables. That is a serious overreach and it will confuse readers.\n\nOther soft spots are more minor but real: no error bars or multiple seeds, hand-set hyperparameters (cluster threshold 0.1, top-50%, K=2) with no sensitivity analysis, and the Moon mAP is explicitly unreliable because the Robbins labels are incomplete. The qualitative results are nice but not load-bearing.\n\nThat said, the paper is not incoherent and the underlying method could be useful once the claims are recalibrated and code/data ship. I would send it to peer review, but with a clear request to fix the headline, add real baselines, and either fix or hedge the Moon experiment. It deserves a serious referee, not a desk reject, but the current version would need major revision before acceptance.","headline":"Useful UDA engineering for space terrain, but the '31% over YOCOv1' headline is contradicted by the paper's own tables and needs recalibration.","tokens_in":15875,"tokens_out":1329,"would_cite":false,"duration_ms":12460,"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":"Adding perceptual consistency and strong feature filtering to similarity-based alignment lets lightweight one-stage YOLO detectors adapt from synthetic to real space terrain without target labels, with mAP gains over 31%.","keywords":["unsupervised domain adaptation","object detection","space terrain","YOLO","visual similarity alignment","perceptual consistency","strong feature filtering","lunar crater detection"],"falsifier":"Re-run the Moon adaptation using a complete hand-labeled crater set, or a subset whose labels are verified complete, and compare PC SFF against ViSGA and Source Only; if the rankings change or the gap collapses, the claimed real-world UDA advantage on lunar data is not established.","tokens_in":1769,"feed_emoji":"🛰️","tokens_out":2594,"duration_ms":77314,"temperature":0.7,"pith_summary":"This paper tries to establish that unsupervised domain adaptation for celestial terrain detection can work in lightweight one-stage detectors, not just in the heavier two-stage detectors where similarity-based alignment originated. The proposed YOCOv2 adds two mechanisms to the existing visual similarity-based alignment scheme: a Perceptual Consistency loss that aligns multi-scale instance features between source and target domains, and Strong Feature Filtering that keeps only the most attention-ranked channels before instance clustering or contrastive matching. On simulated Mars and asteroid data and real lunar mosaics, the method reports mAP improvements of more than 31% over YOCOv1 and terrestrial state-of-the-art baselines, and it runs on a flight-like Zynq/Edge TPU board in real time. If correct, a spacecraft could train on synthetic imagery and then detect craters, dunes, mountains, or boulders on an unexplored body without any manual labels. The paper itself records that the lunar ground-truth catalog is incomplete, so all Moon scores underestimate the true detection ability of every method.","feed_headline":"YOCOv2 lifts space-terrain detection up to 31% without target labels","feed_subtitle":"Two new alignment modules help lightweight YOLO models transfer from synthetic to real Mars, Moon, and asteroid imagery.","key_machinery":"The load-bearing mechanism is the pairing of Perceptual Consistency with Strong Feature Filtering. Perceptual Consistency is a regularization loss that takes the pre-regression instance features produced by YOLO's large, medium, and small detection heads and penalizes their L1 distance between source and target, with weights favoring the smaller spatial resolutions; this disciplines multi-scale feature drift during adversarial alignment. Strong Feature Filtering computes channel attention weights via global average pooling and a one-dimensional convolution, sorts the channels, and keeps the top 50% before instance embeddings are built, so that later clustering and contrastive steps see only the strongest, most terrain-representative channels. The framework also keeps the global adversarial alignment with a gradient reversal layer and explores instance clustering variants and feature clustering variants such as K-Means and pixel top-k attention pooling.","core_discovery":"The central claim of YOCOv2 is that visual similarity-based alignment can be made reliable for one-stage detectors in space terrain by regularizing the pre-regression instance features at all three YOLO scales with an L1 loss weighted toward small scales, and by suppressing weak, ambiguous channels before building instance embeddings so that clustering or contrastive matching groups only perceptually strong terrain features. This combination, called PC SFF, is reported to beat the original ViSGA instance clustering and YOCOv1's intra-feature clustering across most configurations, with the headline gain of nearly 31% observed on Sim Asteroid with YOLO 8-S under contrastive instance clustering (53.0 versus 22.1 mAP@0.5). The paper further claims that the method transfers to real mission imagery, reducing false detections on Mars, asteroid, and lunar scenes, and that the YOLO variants used are fast enough for onboard spacecraft hardware.","pith_inferences":["The headline 'upwards of 31%' rests on a single best-cell comparison on Sim Asteroid with YOLO 8-S under contrastive instance clustering; across the full tables the median advantage of PC SFF over ViSGA is much smaller, so the strength of the central claim depends on how representative that best cell is.","Because the incomplete lunar crater catalog depresses Moon mAP, a fair lunar comparison might actually be more favorable to strong detectors than the tables show, since the paper's own figure suggests predictions contain craters the labels missed.","The channel-ranking idea behind Strong Feature Filtering is architecture-agnostic, so a natural testable extension is to attach PC plus SFF to other one-stage detectors on terrestrial UDA benchmarks and ablate the top-K threshold; the paper only tests a 50% threshold.","The paper's correlation of method family with environment type, where contrastive instance clustering does well on multi-class Mars while feature clustering wins some single-class asteroid cases, suggests that a per-mission VSA selector could be learned or chosen rather than fixed."],"forward_implications":["If the central claim holds, onboard spacecraft can use synthetic-to-real UDA with YOLO nano, small, and medium variants, since these run in under a second on the tested flight-like processor platform and adapt without target labels.","The 31% figure comes from the best configuration on Sim Asteroid, which suggests the method's clearest advantage appears in single-class, texture-repetitive scenes; the multi-class Mars gains are more modest but consistent.","Perceptual Consistency alone, and even pixel top-k attention pooling, outperforms the full PC SFF combination in some architectures, so the paper's own tables indicate the two modules are complementary rather than individually necessary.","On the Moon, all UDA methods land near 33 to 35 mAP while the fully supervised target-only oracle reaches only about 36 to 37 mAP, so the lunar experiment chiefly confirms that a resolution domain gap exists rather than cleanly ranking the UDA methods.","The paper's qualitative results show that source-only training produces many false detections on real mission imagery, and the proposed method reduces those false detections while retaining correct ones, which is the practical outcome that matters for landing and hazard avoidance."],"supporting_citations":[{"why":"This work supplies the ViSGA hierarchical-clustering and contrastive-matching formulation that YOCOv2 adapts to one-stage detectors.","marker":"[12]"},{"why":"This prior YOCO work defines the intra-feature clustering baseline that YOCOv2 extends and compares against.","marker":"[13]"},{"why":"This work supplies the channel attention and pixel top-k pooling mechanisms that Strong Feature Filtering and PTAP build on.","marker":"[48]"},{"why":"This work supplies the gradient reversal layer used for the global adversarial domain alignment.","marker":"[45]"},{"why":"This YOLOv8 reference defines the architecture variant used in the experiment that yields the headline gain.","marker":"[44]"},{"why":"This database supplies the lunar crater locations used as Moon ground truth, and the paper flags it as incomplete.","marker":"[53]"},{"why":"This lunar global mosaic product provides the low- and high-resolution source and target Moon images.","marker":"[52]"},{"why":"This profiling study establishes which YOLO variants run in real time on spaceflight-class processors, supporting the onboard feasibility claim.","marker":"[6]"}],"fun_headline_variants":["YOCOv2: 31% better space terrain detection via strong features","YOCOv2 ups space-terrain UDA by 31% with perceptually consistent features","One-stage UDA for space terrain: YOCOv2 hits 31% gain with PC-SFF","YOCOv2's new alignment cuts reliance on target labels for space detection","Strong features, one-stage: YOCOv2 leads domain-adaptive space terrain detection"],"cache_read_input_tokens":17920,"weakest_assumption_plain":"The lunar evaluation assumes the crater catalog used as ground truth is complete enough to rank detectors, but the paper shows it misses craters that the detectors find, so Moon mAP scores are understated for every method.","fun_headline_variants_meta":{"raw":{"variants":["YOCOv2: 31% better space terrain detection via strong features","YOCOv2 ups space-terrain UDA by 31% with perceptually consistent features","One-stage UDA for space terrain: YOCOv2 hits 31% gain with PC-SFF","YOCOv2's new alignment cuts reliance on target labels for space detection","Strong features, one-stage: YOCOv2 leads domain-adaptive space terrain detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000855,"raw_usage":{"total_tokens":3759,"prompt_tokens":1032,"completion_tokens":2727,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":648,"completion_tokens_details":{"reasoning_tokens":2612}},"tokens_in":648,"tokens_out":2727,"duration_ms":16238,"temperature":1.0,"reasoning_tokens":2612,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:39:49.788878+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Moon adaptation using a complete hand-labeled crater set, or a subset whose labels are verified complete, and compare PC SFF against ViSGA and Source Only; if the rankings change or the gap collapses, the claimed real-world UDA advantage on lunar data is not established.","supporting_citations":[{"cited_title":"Rezaeianaran, R","cited_arxiv_id":null,"evidence_quote":"This work supplies the ViSGA hierarchical-clustering and contrastive-matching formulation that YOCOv2 adapts to one-stage detectors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This work supplies the channel attention and pixel top-k pooling mechanisms that Strong Feature Filtering and PTAP build on."},{"cited_title":"Ganin and V","cited_arxiv_id":null,"evidence_quote":"This work supplies the gradient reversal layer used for the global adversarial domain alignment."},{"cited_title":"Varghese and M","cited_arxiv_id":null,"evidence_quote":"This YOLOv8 reference defines the architecture variant used in the experiment that yields the headline gain."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This database supplies the lunar crater locations used as Moon ground truth, and the paper flags it as incomplete."},{"cited_title":"Speyerer, M","cited_arxiv_id":null,"evidence_quote":"This lunar global mosaic product provides the low- and high-resolution source and target Moon images."},{"cited_title":"Chase Jr, J","cited_arxiv_id":null,"evidence_quote":"This profiling study establishes which YOLO variants run in real time on spaceflight-class processors, supporting the onboard feasibility claim."}],"review_version":1}