{"id":"98e532bd-8c43-47cd-92c6-a95ebf75032a","arxiv_id":"2412.13496","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"QueryCDR introduces learnable query-based control conditions so a single fisheye rectification model can handle multiple distortion degrees without retraining.","lead":"A new neural network uses user-chosen 'queries' to control how much fisheye distortion is corrected in an image. The method reports the best scores on synthetic fisheye benchmarks with nine distortion levels, while real-world tests are only qualitative.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The core generalization claim is untested: Table 1 evaluates only the same nine distortion degrees used for fine-tuning, so QueryCDR's 'without retraining' advantage is not demonstrated for unseen distortion levels.","rationale":"The reader's weakest_assumption focuses on the synthetic-to-real gap, which is a valid concern for practical deployment. However, the more load-bearing issue is that the synthetic evaluation itself does not test generalization to unseen distortion degrees. Sec. 4.1 explicitly uses the same d1–d9 for both fine-tuning and testing, so Table 1 cannot support the claim of handling 'varying degrees of distortion without retraining.' This is an internal inconsistency between the stated contribution and the experiment design. The reader did not flag this specific flaw, though they did note that the test distortion degrees are the same as fine-tuning degrees; in that sense, there is partial agreement. My concern is more fundamental because it invalidates the primary evidence for the central claim even on synthetic data. The concrete test—using held-out degrees with interpolated queries—would directly settle whether the model actually generalizes or merely memorizes. The verdict should remain CONDITIONAL: the paper's architecture is interesting, but the acceptance hinges on demonstrating generalization to unseen distortion degrees, which the current experiments do not provide.","tokens_in":14205,"tokens_out":4155,"duration_ms":35272,"concrete_test":"Evaluate QueryCDR on held-out distortion degrees, e.g., d1.5, d2.5, ..., d8.5, using both interpolated queries (Q_i + 0.5(Q_{i+1} - Q_i)) and nearest-neighbor queries. Report PSNR/SSIM against the same baselines as Table 1. Also test extrapolation with queries outside the d1–d9 range. If performance degrades significantly on held-out degrees, the central generalization claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of QueryCDR is that it generalizes to varying distortion degrees without retraining. However, Sec. 4.1 reveals that the fine-tuning stage uses 18,000 images with distortion degrees d1–d9, and the test set of 3,600 images is drawn from the same nine degrees. Thus, Table 1 measures performance on distortion levels that QueryCDR was explicitly fine-tuned on, not on unseen levels. The model is given a query that encodes each specific degree at test time, so it can memorize a separate condition per degree. This does not demonstrate generalization to a continuum of distortions; it demonstrates conditional generation on nine training conditions. The only evidence for continuous control is qualitative (Fig. 3), with no quantitative results for interpolated or extrapolated queries. Consequently, the claimed superiority over baselines may reflect the unfair advantage of having been fine-tuned on the exact test degrees, while the baselines were not. The generalization claim is therefore internally unsupported by the reported experiments.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes QueryCDR, a fisheye image rectification network that uses a set of learnable queries as controllability conditions. A Distortion-aware Learnable Query Mechanism (DLQM) maps a user-supplied query to layer-wise control conditions, and two modulating blocks (CCMB and CAMB) inject these conditions into a U-shaped rectification network. The model is pre-trained on a single distortion degree and fine-tuned on nine degrees d1–d9, then evaluated on a test set built with the same nine degrees. The authors report large quantitative gains over prior methods on the COCO fisheye benchmark and show qualitative results on the real-world WoodScape dataset. The central claim is that the model can rectify images with varying distortion degrees without retraining, and that interpolating between queries yields continuous control for any degree.","tokens_in":14518,"tokens_out":3472,"duration_ms":33989,"significance":"If the generalization claim were fully supported, this would be a useful step toward practical fisheye rectification, since one model could serve multiple distortion levels and users could control the output by selecting or interpolating queries. The paper is clearly structured, the code is released, and the ablations in Tables 2–4 are systematic and internally consistent. However, the headline claim is currently supported only for the nine distortion levels used in fine-tuning, not for unseen degrees. The real-world validation is exclusively qualitative, and the query-selection protocol for real images is unspecified. As presented, the contribution is a controllable conditional rectification network for nine trained degrees, not a demonstrated generalization to a continuum of distortion degrees. The proposed mechanism is plausible and the reported numbers are strong, but the missing held-out-degree evaluation is load-bearing for the paper's main claim.","major_comments":[{"comment":"The test set uses exactly the nine distortion degrees d1–d9 that were used for fine-tuning (Eq. (13)), and the user supplies a query Qi corresponding to the test degree. Therefore Tab. 1 does not measure generalization to unseen distortion degrees; it measures performance on nine trained conditions. The claims in Sec. 4.2 ('achieves the best performance across all distortion degrees without retraining') and Sec. 3.2 ('smooth continuous rectification for any distortion degree') go beyond what these experiments show. Please add a quantitative evaluation on held-out degrees, e.g., test on d1.5, d2.5, ..., d8.5 using interpolated queries (and possibly extrapolated queries for d0/d10), with PSNR/SSIM and a comparison to the same baselines at those degrees. Without such an experiment, the central generalization claim is unsupported.","section":"Sec. 4.1, Sec. 4.2, Tab. 1"},{"comment":"The control mechanism requires the user to supply a query that corresponds to the distortion degree of the input image. During fine-tuning Eq. (13) explicitly pairs Qi with di, and at test time Tab. 1 uses the same Qi for each di. The paper does not state how a user obtains the correct Qi for a real-world image that was not generated with a known parameter. This makes the 'without retraining' claim operationally dependent on prior knowledge of the distortion degree. Please specify the query-selection protocol for real images, or add an automatic query-prediction module, and report how the queries were chosen for the WoodScape examples in Fig. 5.","section":"Sec. 3.2, Sec. 3.4"},{"comment":"Real-world validation is only qualitative. No quantitative metric is reported on the WoodScape dataset, and there is no comparison against baselines. Since the model is trained entirely on synthetic data from a four-parameter polynomial model, the synthetic-to-real generalization claim requires stronger evidence. At minimum, report a measurable rectification quality score (e.g., line straightness, LPIPS against a corrected reference, or residual distortion parameters) and state the query used for each real image.","section":"Sec. 4.2, Fig. 5"},{"comment":"The comparison with baselines may conflate the benefit of the query mechanism with the benefit of per-degree specialization. The baselines are retrained on all nine degrees jointly but do not receive a per-degree condition, whereas QueryCDR is fine-tuned with a distinct query per degree and can thereby specialize its behavior per degree. The 'W/o Control' row in Tab. 2 replicates the PCN numbers and does not establish what the proposed U-shaped backbone achieves without control. Please add an ablation that trains the QueryCDR backbone on the nine degrees with a single query (or with a scalar degree embedding of comparable capacity) to isolate the contribution of DLQM.","section":"Sec. 4.1, Tab. 1, Tab. 2"}],"minor_comments":[{"comment":"The notation Ql_c is used for both the control condition at layer l and the input to layer l+1; please clarify the dimensions of the FC outputs and state explicitly that when Eq. (3) is used inside layer l, Qc refers to Ql_c.","section":"Sec. 3.2, Eq. (2)"},{"comment":"The interpolation examples Q1.25 = 0.75Q1 + 0.25Q2 and Q8.5 = 0.5Q8 + 0.5Q9 are shown in the figure caption, but the paper does not report any quantitative evaluation of interpolated queries on the COCO test set. Please either add such an evaluation or soften the 'any distortion degree' wording.","section":"Fig. 3 caption"},{"comment":"The experimental settings describe 40,000 pre-training images and 18,000 fine-tuning images, but do not specify how many of these come from COCO versus Places2, and Tab. 1 reports only COCO. Please clarify the dataset split and report Places2 results if they are available.","section":"Sec. 4.1"},{"comment":"In the definition of control-attention, the projection matrices WQ, WK, WV are said to map to R^{m×C}; please specify the relationship between m and the channel dimension C, and whether m equals C in the experiments.","section":"Sec. 3.3, Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable submission to a vision venue, and the query-based control idea is interesting. The main issue is that the key generalization claim is not actually tested: the test distortion degrees coincide with the fine-tuning degrees. I would ask the authors to add held-out-degree experiments and a quantitative real-world evaluation before reconsidering. The code release and thorough ablations are positive factors."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the query-based control idea is real and the paper is worth reading, but the headline generality claim is not supported by the experiments. The core contribution is DLQM, which replaces scalar control with high-dimensional learnable queries and feeds them through CCMB/CAMB blocks into a U-Net. That is a legitimate departure from CFSNet/MM-RealSR, and the ablations show the learned query beats scalar and fixed-query control. The PSNR/SSIM gains in Table 1 are large and consistent across all nine degrees. So there is real engineering value here.\n\nThe soft spot is exactly what the stress-test note says: the nine test distortion degrees are the same nine degrees used in fine-tuning. So Table 1 shows the model can handle nine conditions it was trained for, not that it generalizes to unseen distortion levels. The 'varying degrees without retraining' claim needs an experiment with held-out degrees, e.g., training on d1-d8 and testing on d9, or interpolating queries and measuring against ground truth. The only evidence for continuous control is qualitative (Fig. 3). That is a meaningful gap, not a fatal flaw. The mechanism might well generalize, but the paper doesn't demonstrate it.\n\nAlso, the user must know the distortion degree to pick the right query, which limits the practical 'automatic' story. And real-world validation on WoodScape is only qualitative; no quantitative metrics are reported there. Missing Fishformer from the comparison is a minor citation issue since it is cited but not compared.\n\nFor a CVPR/ICCV-style review, I'd recommend conditional accept with requests for held-out-degree evaluation and real-world quantitative results. The ablations are honest and the architecture is sensible. It deserves a serious referee.","headline":"A genuinely new control mechanism for fisheye rectification, but the generalization claim is only tested on the nine fine-tuned distortion levels, not on unseen ones.","tokens_in":14941,"tokens_out":1562,"would_cite":true,"duration_ms":13952,"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":"QueryCDR claims that a single network, guided by user-selectable learnable queries, rectifies fisheye images at every tested distortion degree without retraining.","keywords":["fisheye image rectification","controllable rectification","learnable queries","distortion-aware control","appearance flow","CNN-Transformer hybrid","generalization without retraining","image restoration"],"falsifier":"On real fisheye images from lenses with different fields of view, pick the nearest trained query, rectify, and measure the residual curvature of straight scene lines or compare against calibration-based undistortion. If the rectified outputs keep visible line curvature at distortion levels away from the nine anchors, or if the quality gap to a per-degree retrained model grows with distance from the anchors, the central no-retraining claim would be refuted.","tokens_in":13993,"feed_emoji":"📷","tokens_out":9825,"duration_ms":81305,"temperature":0.7,"pith_summary":"The paper sets out to establish that one fisheye-rectification network can handle the full range of distortion strengths without being retrained for each lens. The design replaces a single scalar degradation level with learnable spatial queries: the user selects or interpolates a query, and the network modulates its own features at every layer according to that query. On a nine-level synthetic benchmark built from COCO, QueryCDR reports the best image-quality metrics PSNR and SSIM at every level, averaging 20.32 dB against 18.53 dB for the previous best method. If this holds, fisheye cameras with different fields of view would no longer force retraining or a separate model per distortion degree.","feed_headline":"One model fixes fisheye images at every tested distortion level","feed_subtitle":"Query-based control beats prior best by 1.8 dB on COCO and needs no retraining between degrees.","key_machinery":"The central object is the Distortion-aware Learnable Query Mechanism (DLQM): a set of learnable queries, each the same spatial size as the input image, encoding position-dependent distortion patterns instead of a single number. DLQM extracts a query into per-layer control conditions through convolutions and fully connected layers, and two modulating blocks apply them: the Controllable Convolution Modulating Block (CCMB) predicts a dynamic fusion ratio between original and controlled features, while the Controllable Attention Modulating Block (CAMB) uses the controlled feature as the attention query to capture long-range distortion mappings. Together they let one network adjust its rectification strength purely from the query input.","core_discovery":"On its own terms, the discovery is that distortion rectification can be controlled by position-dependent latent queries rather than a scalar degradation index. During training, the nine learnable queries in the Distortion-aware Learnable Query Mechanism (DLQM) each absorb the latent spatial mapping of one distortion degree; at inference, the chosen query is converted layer by layer into control conditions that steer both local convolution modulation and global attention modulation. The paper's evidence is that this design outperforms every compared method at every distortion level on the COCO fisheye benchmark, and that interpolating between queries produces smooth intermediate rectifications.","pith_inferences":["If the query set forms a continuous manifold, a natural next step is to regress the query from the image itself, making rectification fully automatic rather than user-selected.","The same query-conditioned modulation could apply to other spatially structured degradations, such as radial vignetting, non-uniform motion blur, or panoramic projections, where scalar controls are too coarse.","Measuring line straightness on WoodScape rectifications would quantitatively test whether the synthetic four-parameter distortion family actually covers real lenses; the paper only offers qualitative evidence."],"forward_implications":["A single trained model can replace per-distortion retraining: switching from mild to severe distortion only requires selecting a different query.","Query interpolation gives continuous control, so users can produce rectifications at distortion strengths between the trained anchors.","The hybrid CNN/Transformer architecture trades off quality and cost, with the 6C+5A configuration giving the best measured balance of PSNR, SSIM, FLOPs, and parameter count.","Ablations show that spatial query conditions outperform scalar or fixed-position controls, indicating that position-dependent control is what enables the generalization.","Qualitative results on WoodScape indicate that the synthetic-trained queries carry over to real fisheye images, although the paper reports no quantitative real-world metric."],"supporting_citations":[{"why":"Supplies the appearance-flow warping and U-shaped architecture that QueryCDR builds on, and provides the uncontrollable baseline in the ablation study.","marker":"[52]"},{"why":"Gives the previous best rectification method, named SimFIR, which QueryCDR outperforms by 1.79 dB average PSNR in the main comparison.","marker":"[10]"},{"why":"Provides the dual-diffusion baseline DDA, the strongest prior method at several distortion levels in Table 1.","marker":"[54]"},{"why":"Provides the GAN-based DR-GAN baseline used in both quantitative and qualitative comparisons.","marker":"[24]"},{"why":"Supplies the COCO source images used to build the synthetic fisheye training and test sets for all quantitative experiments.","marker":"[27]"},{"why":"Supplies the Places2 source images used as a second dataset for constructing synthetic fisheye training data.","marker":"[62]"},{"why":"Supplies the WoodScape real-world fisheye images used to qualitatively support generalization from synthetic training.","marker":"[57]"}],"fun_headline_variants":["Fish-eye fix: Query control adapts to any distortion","Warped no more: Query-based model fixes fish-eye at any distortion","Learnable queries let fish-eye model adapt to any distortion level","One fish-eye fixer for all distortions: QueryCDR beats prior best","No retraining: QueryCDR fixes fish-eye at any distortion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the four-parameter polynomial model used to synthesize training images faithfully represents the distortion of real fisheye cameras, so the query-conditioned mappings learned on synthetic data remain valid in practice; the paper validates this only qualitatively on WoodScape.","fun_headline_variants_meta":{"raw":{"variants":["Fish-eye fix: Query control adapts to any distortion","Warped no more: Query-based model fixes fish-eye at any distortion","Learnable queries let fish-eye model adapt to any distortion level","One fish-eye fixer for all distortions: QueryCDR beats prior best","No retraining: QueryCDR fixes fish-eye at any distortion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000962,"raw_usage":{"total_tokens":4072,"prompt_tokens":893,"completion_tokens":3179,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":3087}},"tokens_in":509,"tokens_out":3179,"duration_ms":20443,"temperature":1.0,"reasoning_tokens":3087,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:04:12.161098+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On real fisheye images from lenses with different fields of view, pick the nearest trained query, rectify, and measure the residual curvature of straight scene lines or compare against calibration-based undistortion. If the rectified outputs keep visible line curvature at distortion levels away from the nine anchors, or if the quality gap to a per-degree retrained model grows with distance from the anchors, the central no-retraining claim would be refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the appearance-flow warping and U-shaped architecture that QueryCDR builds on, and provides the uncontrollable baseline in the ablation study."},{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Gives the previous best rectification method, named SimFIR, which QueryCDR outperforms by 1.79 dB average PSNR in the main comparison."},{"cited_title":"Dual Diffusion Architecture for Fisheye Image Rectification: Synthetic-to-Real Generalization","cited_arxiv_id":"2301.11785","evidence_quote":"Provides the dual-diffusion baseline DDA, the strongest prior method at several distortion levels in Table 1."},{"cited_title":"IEEE Transactions on Circuits and Systems for Video Technology30(3), 725–733 (2019)","cited_arxiv_id":null,"evidence_quote":"Provides the GAN-based DR-GAN baseline used in both quantitative and qualitative comparisons."},{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Supplies the WoodScape real-world fisheye images used to qualitatively support generalization from synthetic training."}],"review_version":1}