{"id":"7be5c209-2200-4249-a481-d9998f0f92cd","arxiv_id":"2606.22834","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Homographic navigation uses homography-driven synthetic data generation and a two-pass keypoint model with Stable Warp training to achieve accurate planar alignment from minimal supervision.","lead":"This paper introduces homographic navigation, a framework that treats homography as the central organizing variable to train a keypoint model from one reference image for precise planar region capture. A smart generalist might read it to see how geometric augmentation can reduce labeling needs in camera guidance tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Synthetic homographies from one reference image may not capture real-world photometric and geometric variations needed for generalization.","rationale":"The reader's weakest_assumption directly identifies the same generalization bottleneck. With full text now accessible, the concern remains load-bearing because the method's novelty (homography as organizing variable, Stable Warp) does not automatically resolve the domain gap; only an explicit real-vs-synthetic ablation would settle it. No other internal inconsistency appears from the abstract and stated claims.","tokens_in":1644,"tokens_out":363,"duration_ms":11960,"concrete_test":"Split the evaluation into (a) synthetic test warps drawn from the same reference and (b) a new set of real photographs of the same physical artifacts taken with varied cameras, lighting, and slight non-planar bends. Report keypoint reprojection error and final alignment success rate on both; if real-world error exceeds synthetic error by >30% or success rate falls below 80%, the minimal-supervision claim does not transfer.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that a model trained solely via homographic warps of a single annotated reference image (plus the two-pass inference and Stable Warp strategy) produces accurate keypoint predictions on real camera feeds. This rests on the untested premise that 2D projective augmentations alone suffice to simulate the joint distribution of lighting, sensor noise, lens distortion, partial occlusions, and non-rigid surface deviations encountered in the wild. If the experimental section only reports metrics on held-out synthetic warps (or on the same reference image family), the generalization step remains unsupported; the abstract's mention of \"future learning from in-the-wild video data\" acknowledges this gap rather than closing it.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper presents homographic navigation, a geometry-centric framework that treats homography as an organizing variable for guiding camera capture of planar regions. From a single annotated reference image, unlimited synthetic training data is generated via homographic augmentation to train a single-shot model for joint recognition and localization of rectangular planar artifacts via sparse keypoint prediction. A two-pass inference scheme (global detection followed by localized refinement) and Stable Warp training strategy are introduced to improve precision under limited input resolution, along with per-keypoint and per-sample confidence estimates. The central claim is that accurate planar alignment can be achieved from minimal supervision, providing a foundation for geometry-driven camera guidance and future in-the-wild learning.","tokens_in":1780,"tokens_out":646,"duration_ms":16256,"significance":"If the generalization claim holds, the approach would be significant for enabling high-precision geometric tasks with extremely low supervision (one reference image), reducing reliance on large annotated datasets in computer vision. The unification of learning, alignment, and evaluation around homography is a conceptual strength, and the two-pass scheme plus Stable Warp strategy address a practical resolution-precision tradeoff. However, the significance is conditional on evidence that synthetic 2D projective augmentations suffice for real-world generalization.","major_comments":[{"comment":"Experiments section: all reported metrics and ablations (including improvements from Stable Warp) are confined to held-out synthetic homographic warps of the reference image; no quantitative results, baselines, or error analysis on real camera captures (varying illumination, lens distortion, partial occlusion, or non-rigid surfaces) are provided, which is load-bearing for the generalization premise stated in the abstract and §1.","section":"Experiments"},{"comment":"§3.2 (two-pass inference) and §3.3 (Stable Warp): the refinement pass operates on cropped regions from the first pass, but the training distribution remains purely synthetic; without a real-world test set or domain-adaptation experiment, the claim that this yields 'accurate planar alignment' on physical camera feeds is unsupported.","section":"§3.2 and §3.3"}],"minor_comments":[{"comment":"Abstract: the phrase 'experimental results demonstrate' should be accompanied by at least one concrete metric or dataset descriptor to avoid overstatement.","section":"Abstract"},{"comment":"Notation: the per-keypoint and per-sample confidence outputs are mentioned but lack an explicit formulation or loss term; adding this in §3 would improve clarity.","section":"§3"},{"comment":"Figure captions: several figures showing keypoint predictions would benefit from explicit scale bars or pixel-error annotations to aid interpretation of the high-precision regime.","section":null}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonable fit for cs.CV but the absence of any real-world validation data makes the central claim difficult to assess; the authors acknowledge the gap via the 'future learning from in-the-wild video data' remark, which suggests the current experiments are preliminary."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback. We address each major comment below, acknowledging the synthetic scope of the current experiments and committing to revisions that clarify the claims without overstating generalization.","responses":[{"response":"We agree that all reported metrics and ablations are performed exclusively on synthetic homographic warps of the single reference image. This is intentional, as the framework's central contribution is the use of homography as an organizing variable to generate unlimited training data from minimal (one-image) supervision. The experiments rigorously validate the two-pass inference and Stable Warp strategy within this controlled synthetic regime, demonstrating precision gains. However, the abstract and §1 do imply broader applicability. We will revise the abstract, introduction, and conclusion to explicitly qualify that accurate planar alignment is shown in synthetic evaluations, and that extension to real camera captures (with factors such as illumination variation or occlusion) constitutes future work. A limitations section will be added to discuss the current synthetic-to-real gap.","revision_made":"yes","referee_comment":"[Experiments] Experiments section: all reported metrics and ablations (including improvements from Stable Warp) are confined to held-out synthetic homographic warps of the reference image; no quantitative results, baselines, or error analysis on real camera captures (varying illumination, lens distortion, partial occlusion, or non-rigid surfaces) are provided, which is load-bearing for the generalization premise stated in the abstract and §1."},{"response":"The training distribution is purely synthetic, as the method generates data via homographic augmentation from one annotated reference; no real images are used for training or the reported ablations. The two-pass scheme and Stable Warp are evaluated synthetically to isolate their effect on precision at limited resolution. We concur that unsupported claims about physical camera feeds should be avoided. We will revise §3.2, §3.3, the abstract, and related text to state that the components yield accurate alignment under synthetic conditions and provide a foundation for future real-world deployment, rather than asserting direct applicability to in-the-wild physical feeds.","revision_made":"yes","referee_comment":"[§3.2 and §3.3] §3.2 (two-pass inference) and §3.3 (Stable Warp): the refinement pass operates on cropped regions from the first pass, but the training distribution remains purely synthetic; without a real-world test set or domain-adaptation experiment, the claim that this yields 'accurate planar alignment' on physical camera feeds is unsupported."}],"tokens_in":1385,"tokens_out":537,"duration_ms":21638,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core contribution is a framework that treats homography as the central variable: start with one annotated reference, generate unlimited synthetic warps for training, then run a single-shot keypoint predictor with a two-pass inference step for refinement and a Stable Warp training strategy that improves high-precision results. It also outputs per-keypoint and per-sample . This setup aims at guiding a camera to capture planar regions accurately with minimal labeling.\n\nWhat stands out is the particular combination and the Stable Warp component, which the authors say helps noticeably in the precision regime. The geometry-driven angle is sensible and directly addresses the labeling bottleneck for tasks like artifact localization.\n\nThe main limitation is generalization. Training exclusively on projective warps of a single image leaves out photometric variation, lens effects, partial occlusions, and any deviation from perfect planarity. The abstract flags future work on in-the-wild video, which signals that current results likely stay within synthetic or closely matched test distributions. If the experiments do not include diverse real captures with quantitative comparison to baselines, the accuracy claims rest on an untested assumption.\n\nThe underlying math is standard homography and keypoint regression, with no obvious inconsistencies. Citation patterns look normal for this sub-area.\n\nThis is worth a serious referee for groups working on low-supervision planar capture or camera guidance pipelines. The idea is practical and the method is reproducible in principle, so it deserves review even if more real-world validation is needed.","headline":"The paper gives a clean way to train keypoint models for planar alignment from one reference image using synthetic homographies, plus a two-pass scheme and Stable Warp trick, but real-world generalization is the weak link.","tokens_in":2287,"tokens_out":381,"would_cite":false,"duration_ms":21294,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A model trained solely on synthetic homographic augmentations from one reference image can guide cameras to precise planar alignment via keypoint prediction and two-pass refinement.","keywords":["homographic navigation","planar alignment","keypoint prediction","synthetic augmentation","two-pass inference","camera guidance","minimal supervision","Stable Warp training"],"falsifier":"Test the trained model on a set of real camera captures of the target planar objects under lighting, distances, or angles outside the augmentation range; if average alignment error exceeds the claimed precision threshold, the approach does not generalize as stated.","tokens_in":2553,"feed_emoji":"📐","tokens_out":720,"duration_ms":15608,"temperature":0.7,"pith_summary":"The paper establishes homographic navigation as a framework that treats homography as the central organizing variable for learning and camera guidance toward exact planar captures. From a single annotated reference, unlimited synthetic training data is created by applying homographic transformations, allowing a single-shot model to jointly recognize and localize rectangular planar artifacts through sparse keypoint prediction. A two-pass inference process first detects globally then refines locally, paired with a Stable Warp training approach, to reach high precision even at limited input resolutions while also outputting per-keypoint and per-sample confidence scores. This setup demonstrates that accurate alignment is possible with minimal real supervision rather than large labeled datasets.","feed_headline":"One reference image trains model for exact planar camera captures","feed_subtitle":"Homographic augmentations plus two-pass refinement reach high-precision alignment without large real datasets.","key_machinery":"Homographic navigation, the geometry-centric framework that treats homography as the unifying variable across data generation, model training, and inference for planar alignment.","core_discovery":"Homographic navigation unifies learning, alignment, and evaluation by using homography as the organizing variable. Synthetic homographic augmentations generated from one reference image train a model for sparse keypoint prediction on multiple planar artifacts. The two-pass inference scheme performs global detection followed by localized refinement, and the Stable Warp training strategy boosts accuracy in the high-precision regime, with the model also estimating confidence for each keypoint and the full sample.","pith_inferences":["The same synthetic-augmentation idea could be tested on non-rectangular planar targets by adjusting the keypoint layout.","Integrating the confidence scores into an active feedback loop might let a camera app automatically adjust position until alignment confidence exceeds a threshold.","If the two-pass refinement step is replaced by a higher-resolution single pass, the accuracy gain from Stable Warp could be measured directly against that baseline.","The approach may apply to other geometry-driven tasks such as document scanning or surface inspection where exact planar registration is required."],"forward_implications":["Accurate planar alignment becomes feasible with only one annotated reference image instead of large real-world datasets.","The two-pass scheme combined with Stable Warp training delivers measurable gains in high-precision keypoint localization.","Per-keypoint and per-sample confidence predictions are produced alongside the alignment output.","The method supplies a starting point for geometry-driven camera guidance systems.","It opens a path to training directly from unlabeled in-the-wild video sequences."],"fun_headline_variants":["Homographic navigation unifies geometry for planar camera guidance","Single reference generates unlimited synthetic training via homography","Sparse keypoint model trained with Stable Warp for planar artifacts","Two-pass scheme refines global detections for planar alignment"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Synthetic homographic changes applied to one reference image create enough variety and realism for the trained model to work on actual camera images under different conditions.","fun_headline_variants_meta":{"raw":{"variants":["Homographic navigation unifies geometry for planar camera guidance","Single reference generates unlimited synthetic training via homography","Sparse keypoint model trained with Stable Warp for planar artifacts","Two-pass scheme refines global detections for planar alignment"]},"model":"grok-4.3","cost_usd":0.004099,"raw_usage":{"total_tokens":2053,"prompt_tokens":612,"num_sources_used":0,"completion_tokens":61,"cost_in_usd_ticks":40987000,"prompt_tokens_details":{"text_tokens":612,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1380,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":612,"tokens_out":61,"duration_ms":10875,"temperature":1.0,"reasoning_tokens":1380,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T09:33:05.143260+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Test the trained model on a set of real camera captures of the target planar objects under lighting, distances, or angles outside the augmentation range; if average alignment error exceeds the claimed precision threshold, the approach does not generalize as stated.","supporting_citations":[],"review_version":1}