{"id":"c90b6429-e075-4334-887a-3e2123762867","arxiv_id":"1908.06837","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A cGAN-based image de-fencing system, including a faster single-stage variant conditioned on Canny edges, is proposed and shown qualitatively on synthetic fence images.","lead":"The paper applies conditional GANs to remove fence patterns from photos, using a two-stage mask-and-inpaint pipeline and a faster single-stage network that also receives Canny edge maps as input. It claims the single-stage system matches the two-stage visual quality at lower runtime, but the preprint presents only qualitative examples and no numeric evaluation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Two-stage baseline is not well-specified: Eqs. (1) and (5) condition generators on ground-truth targets, so the test-time comparison in Section 4.3 cannot be reproduced as written.","rationale":"My read agrees with the reader that the paper lacks quantitative support; the absence of tables, baselines, and code is real. However, I think the more load-bearing issue is that the two-stage network, which is the explicit comparator for the central claim, is underspecified: Eq. (1) gives G1 the ground-truth mask and Eq. (5) gives G2 the ground-truth de-fenced image as inputs, while at test time these are unavailable. This could be a notation error, but the manuscript never states the corrected conditioning, and no code is provided to disambiguate. The claim that the single-stage network produces 'similar visual quality' to the two-stage network cannot be assessed if the two-stage protocol is not defined. The Canny-edge-conditioned single-stage idea is plausible and the qualitative figures are suggestive, but the manuscript's stated evidence is insufficient. Therefore the existing CONDITIONAL verdict remains appropriate; no change in disposition.","tokens_in":9559,"tokens_out":6979,"duration_ms":73532,"concrete_test":"Obtain the authors' implementation or ask them for the exact test-time forward pass. A specific check: run the two-stage pipeline on the 245-image test set using only fenced images. If G1 or G2 requires Imask or Idef at inference, the baseline is unimplementable; if not, verify that the training objective matches Eq. (12) and report PSNR/SSIM for single-stage versus two-stage. Without this, the equivalence claim is unverified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparison in Section 4.3 is between the single-stage network and the two-stage network of Section 3.1. But the two-stage network is not well-specified at inference time. Eq. (1) defines the mask generator as Ipred = G1(Ifen, Imask), i.e., it takes the ground-truth fence mask as an input; Eq. (5) defines the recovery network as \\tilde{I}pred = G2(\\tilde{I}fen, Idef), taking the ground-truth de-fenced image as an input. The Abstract states that at evaluation the fenced image is passed through the two generators in succession. If Eqs. (1) and (5) are literal, Imask and Idef are unavailable during evaluation, so the two-stage pipeline cannot run as described; if they are not literal, the actual test-time inputs and training losses are never stated, and the 'similar visual quality' claim rests on an unspecified baseline. The manuscript also claims quantitative results and public code/data, but Section 4 contains only sample images and no metrics, tables, baselines, or repository links. The strongest claim therefore lacks both a reproducible comparison and a defined baseline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two cGAN-based image de-fencing pipelines: a two-stage network that first predicts a fence mask and then inpaints the masked regions, and a single-stage network that directly maps a fenced image plus its Canny edge map to a de-fenced image. The authors claim that the single-stage network achieves visual quality comparable to the two-stage network while being faster, and that comparative experiments show superiority over existing de-fencing methods. The manuscript describes the network losses in detail but the experimental section (Section 4) contains only qualitative image examples; no numeric metrics, baseline comparisons, or reproducibility artifacts are provided.","tokens_in":9792,"tokens_out":4200,"duration_ms":44408,"significance":"If supported, the single-stage design would be a practically useful contribution because it targets the latency bottleneck of sequential mask-and-inpaint pipelines and uses a simple edge prior. The related-work survey is thorough and the loss design follows established cGAN inpainting practice. However, as submitted the paper's central empirical claims are unverifiable: the abstract promises quantitative results and comparisons, but Section 4 presents none, the two-stage baseline is not reproducible as written, and the single-stage evaluation rests on a small set of qualitative examples. The contribution therefore cannot be assessed beyond a plausible architectural proposal.","major_comments":[{"comment":"The abstract states that 'qualitative and quantitative experimental results reported in the manuscript' support the claims, and Section 4.3 states a 'comparative performance analysis' with state-of-the-art techniques, but no quantitative results appear anywhere. There are no PSNR, SSIM, or any other numeric metrics, no tables, no error bars, and no comparison to prior de-fencing methods. The central claim of comparable quality and superiority over existing methods is therefore unsupported by any evidence.","section":"Abstract and Section 4"},{"comment":"The two-stage pipeline is not reproducible as written. Eq. (1) defines the mask generator as Ipred = G1(Ifen, Imask), conditioning on the ground-truth fence mask, and Eq. (5) defines the recovery network as \\tilde{I}pred = G2(\\tilde{I}fen, Idef), conditioning on the ground-truth de-fenced image. Since these ground-truths are unavailable at evaluation, the test-time comparison in Section 4.3 between the single- and two-stage networks cannot be executed as described. The paper must state clearly which inputs are used at train time versus test time and rewrite the equations accordingly; if conditioning on ground truth is intentional, the task becomes trivial for the generator.","section":"Section 3.1.1, Eq. (1) and Section 3.1.2, Eq. (5)"},{"comment":"The synthetic fence generation procedure is not specified. The text says only that 'artificial fence structures' are added to Pascal VOC and COCO images, but gives no details of fence types, densities, orientations, or overlay protocol. This makes the dataset irreproducible and prevents assessment of whether the synthetic distribution is representative of real fences. Moreover, because both training and test sets are generated by the same overlay process, the reported qualitative success partly reflects the network's ability to invert that specific synthetic operation rather than general de-fencing ability; the few real images in the test set have no ground truth and are not evaluated quantitatively.","section":"Section 4.1"},{"comment":"The claim that the single-stage network achieves 'similar visual quality' to the two-stage network with 'fewer number of layers' is not supported by any quantitative comparison or by reporting the actual number of layers or inference time of either network. The comparison in Figure 5 is purely visual, on a small selection of images, and no evidence is given that the single-stage network is indeed faster in terms of parameters, FLOPs, or wall-clock time. Without such measurements, the central efficiency claim is unsubstantiated.","section":"Section 4.3 and Section 3.2"},{"comment":"The contribution of the Canny edge map channel is not ablated. Although Figure 5 shows outputs with and without Canny supervision on a few images, the paper provides no quantitative measure (e.g., PSNR/SSIM on the test set) of how much the edge map improves reconstruction, and no analysis of its robustness to the Canny detector's threshold parameters. Since the paper's central proposal is the use of edge maps to compensate for fewer layers, this missing ablation is a load-bearing gap.","section":"Section 3.2"}],"minor_comments":[{"comment":"There is a mismatched parenthesis in Eq. (2); the expression should be read as min_G1 (α1 max_D1 Ladv,1) + β1 LL1,1. Please correct the typesetting.","section":"Section 3.1.1, Eq. (2)"},{"comment":"The SSIM formula in Eq. (11) is garbled; the terms 'µx2µy 2 +C1' and 'σx2σy 2 +C2' should be rendered with proper superscripts. As written, the definition is unreadable.","section":"Section 3.1.2, Eq. (11)"},{"comment":"The Canny edge filter is used without a citation. Please add the classic reference (Canny, 1986).","section":"References"},{"comment":"In the reference to Zheng and Kambhamettu, 'Digital mmtting' should be 'Digital matting'.","section":"References"},{"comment":"The contributions state that the codes and dataset are publicly available, but no repository link or dataset URL is provided anywhere in the manuscript. This should be added or the claim removed.","section":"Contributions and Section 4.1"}],"recommendation":"reject","confidential_remarks":"This manuscript appears to be an early draft rather than a complete submission. The absence of any quantitative evaluation despite explicit claims of quantitative results, the ill-specified two-stage baseline in Eqs. (1) and (5), and the lack of a reproducible synthetic data protocol are fundamental issues that would require a major experimental overhaul. I do not see a path to acceptance within the normal revision cycle unless the authors essentially rewrite the experimental section and clarify the training/test formulation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You asked about arXiv:1908.06837. Quick take: the paper has a sensible idea, but as submitted it does not support the claims made in its abstract. If the authors supply the missing numbers and fix a real specification problem, it could become a useful incremental contribution.\n\nWhat's new: this seems to be the first cGAN-based end-to-end de-fencing work, at least among the papers it cites. The single-stage variant, where the generator gets the fenced image plus a Canny edge map, is a small but legitimate twist on Pix2Pix/EdgeConnect. The writing is clear, and the qualitative figures show plausible outputs.\n\nThe problems are mostly evidentiary, and they are load-bearing. The abstract promises \"quantitative experimental results\" and a comparison with state-of-the-art, but Section 4 has only eight sample images. No PSNR/SSIM, no tables, no baseline numbers, no timing measurements, even though the whole motivation for the single-stage model is speed. The paper also says codes and data will be released, but gives no links.\n\nThere is also a more technical issue that the stress-test note gets right. Equations (1) and (5) define the generators as taking the ground-truth mask and the ground-truth de-fenced image as inputs—G1(Ifen, Imask) and G2(Ifen~, Idef). If those equations are literal, the two-stage baseline can't be run at test time, because the ground-truths aren't available. If they're shorthand for the standard Pix2Pix conditioning, then the actual inference inputs are never stated. Either way, the central comparison against the two-stage network rests on an unspecified baseline. That is a serious reproducibility problem, not a cosmetic one.\n\nThe evaluation design is also circular in an ordinary sense: fences are synthetically overlaid on clean images, and the test set comes from the same procedure. That doesn't make the method wrong, but it does mean the reported visual quality may not transfer to real fence photos.\n\nIs there anything good beyond the idea? The loss design is standard and the parameter choices are explicit. The authors honestly note at the end that irregular or repeated fence structures remain a challenge. Minor positives, but real.\n\nBottom line: this is not a paper that should be accepted as is. It is also not a waste of a referee's time, because the core idea is plausible and the required corrections are clearly scoped: add actual metrics, specify the inference-time inputs, run an ablation on the Canny channel, and release the promised artifacts. I would send it to review with an expectation of major revision. For your own work, I wouldn't cite the numbers, but I might mention it as evidence that GAN-based de-fencing exists.\n\nRecommendation: engage with it as a revise-and-resubmit candidate, not as a finished contribution.","headline":"A plausible extension-level idea for GAN-based de-fencing that is currently unsupported by missing quantitative evidence and an underspecified two-stage baseline; worth a serious referee only as a major-revision candidate.","tokens_in":10368,"tokens_out":3064,"would_cite":false,"duration_ms":29336,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single conditional GAN, conditioned on a Canny edge map of the fenced photo, can de-fence images with visual quality close to that of a two-network mask-and-inpaint pipeline while needing fewer layers and less time.","keywords":["image de-fencing","conditional generative adversarial networks","Canny edge map","fence mask generation","image inpainting","single-stage image translation","edge-conditioned generation"],"falsifier":"Take a set of real scenes photographed both through a fence and from the same viewpoint without the fence, run both the single-stage and two-stage networks, and compare the de-fenced output with the unobstructed photo; if the single-stage output is visibly worse or scores clearly lower on structural similarity, the paper's parity claim fails.","tokens_in":9335,"feed_emoji":"🖼️","tokens_out":8192,"duration_ms":74148,"temperature":0.7,"pith_summary":"Fences and cage bars in photos block the view of the intended subject, and removing them usually takes two separate steps: find the fence pixels, then fill in what they hide. This paper tries to show that both steps can be replaced by a single conditional generative adversarial network (cGAN) that maps a fenced image directly to a de-fenced image. To keep the one-network version accurate despite having fewer layers, the generator is trained not only on fenced/de-fenced image pairs but also on a Canny edge map of the fenced input, which tells it where the fence lines are. The authors report that this single-stage network produces de-fenced images with visual quality similar to their own two-stage mask-then-inpaint pipeline while being faster, and that both GAN-based approaches beat earlier semi-automated de-fencing techniques. If that holds, fast and fully automatic fence removal becomes practical for ordinary photographs.","feed_headline":"One cGAN removes fences from photos as well as a two-stage network","feed_subtitle":"Single-stage de-fencing with Canny edge guidance reaches two-stage visual quality in fewer layers.","key_machinery":"The central mechanism is the edge-conditioned input to a single cGAN generator: the fenced RGB image and its Canny edge map are stacked and passed through an encoder–decoder, while a PatchGAN discriminator—a network that classifies small 16×16 image patches as real or fake—provides the adversarial feedback. The Canny map is a cheap hand-crafted prior that makes the periodic fence lines explicit, so the generator can locate occluding structure even without the extra layers of the two-stage pipeline. Training minimizes a composite objective: adversarial loss plus L1, perceptual, style, and SSIM losses, mirroring the loss of the two-stage image-recovery network.","core_discovery":"The paper argues that image de-fencing can be compressed into a single generator. Its two-stage baseline first runs a cGAN that outputs a fence mask from a fenced image, then a second cGAN that inpaints the masked image; the single-stage variant skips the mask step and instead concatenates a Canny edge map of the fenced image with the RGB input, feeding the pair to one encoder-decoder generator and one PatchGAN discriminator. The paper reports that this single-stage network produces de-fenced images visually similar to the two-stage outputs, despite having fewer layers, and that both GAN-based approaches improve on earlier de-fencing methods. It identifies the Canny edge map as the ingredient that preserves quality in the shallower network.","pith_inferences":["[Editorial inference] Because the edge map supplies the periodic structure, the same input trick likely transfers to other repeating occluders, such as window grilles, chain-link mesh, or slatted blinds, after retraining.","[Editorial inference] The training data are synthetic fence overlays on general photos, so the parity claim is probably strongest for regular, near-textureless fence patterns; real fences with irregular spacing or strong cast shadows may widen the gap unless real paired data are added.","[Editorial inference] If the single-stage speed holds up, adding a temporal-consistency term could turn this into a practical video de-fencing method, which the paper lists as future work."],"forward_implications":["A single forward pass through one generator can replace the two-network mask-and-inpaint sequence, so de-fencing is faster and needs less memory.","No user interaction is required at test time: the edge map is computed automatically and the same generator handles any input of the trained size.","If the quality comparison holds, practitioners can choose the single-stage model whenever response time matters without giving up much visual quality.","The two GAN-based formulations give a common learning-based baseline that later de-fencing work can be measured against."],"supporting_citations":[{"why":"Supplies the cGAN encoder–decoder generator and PatchGAN discriminator architecture used by both de-fencing networks.","marker":"Isola et al. [2017]"},{"why":"Provides the two-stage adversarial edge-learning inpainting design and composite loss that the image-recovering network adapts.","marker":"Nazeri et al. [2019]"},{"why":"Defines the perceptual loss used to compare high-level features of generated and ground-truth de-fenced images.","marker":"Johnson et al. [2016]"},{"why":"Defines the style loss based on Gram matrices used in the objective.","marker":"Gatys et al. [2016]"},{"why":"Defines the SSIM index used as a structural-similarity loss term.","marker":"Wang et al. [2004]"},{"why":"Supplies loss-function analysis and the SSIM-based loss formulation used in training.","marker":"Zhao et al. [2017]"},{"why":"Provides the public fence-segmentation dataset used to train the fence-mask generator and test data.","marker":"Du et al. [2018]"},{"why":"Provides Pascal VOC images used to create synthetic fenced training and test examples.","marker":"Everingham et al. [2010]"},{"why":"Provides COCO images used to construct the synthetic de-fencing datasets.","marker":"Lin et al. [2014]"}],"fun_headline_variants":["Single cGAN speeds up fence removal, matches two-stage quality","Canny edge map lets one cGAN rival two-stage de-fencing","Fewer layers, same de-fencing: edge-guided cGAN wins","One generator, Canny edges: faster de-fencing without quality drop"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the synthetic fences overlaid on ordinary dataset images during training are representative enough of real fences that the reported visual-parity results transfer to actual fenced photographs.","fun_headline_variants_meta":{"raw":{"variants":["Single cGAN speeds up fence removal, matches two-stage quality","Canny edge map lets one cGAN rival two-stage de-fencing","Fewer layers, same de-fencing: edge-guided cGAN wins","One generator, Canny edges: faster de-fencing without quality drop"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000288,"raw_usage":{"total_tokens":1722,"prompt_tokens":1011,"completion_tokens":711,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":634}},"tokens_in":627,"tokens_out":711,"duration_ms":6782,"temperature":1.0,"reasoning_tokens":634,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:32:49.955679+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of real scenes photographed both through a fence and from the same viewpoint without the fence, run both the single-stage and two-stage networks, and compare the de-fenced output with the unobstructed photo; if the single-stage output is visibly worse or scores clearly lower on structural similarity, the paper's parity claim fails.","supporting_citations":[{"cited_title":"Image style transfer using convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"Defines the style loss based on Gram matrices used in the objective."},{"cited_title":"Image quality assessment: from error visibility to structural similarity","cited_arxiv_id":null,"evidence_quote":"Defines the SSIM index used as a structural-similarity loss term."},{"cited_title":"Loss functions for image restoration with neural networks","cited_arxiv_id":null,"evidence_quote":"Supplies loss-function analysis and the SSIM-based loss formulation used in training."},{"cited_title":"The pascal visual object classes (voc) challenge","cited_arxiv_id":null,"evidence_quote":"Provides Pascal VOC images used to create synthetic fenced training and test examples."},{"cited_title":"Microsoft coco: Common objects in context","cited_arxiv_id":null,"evidence_quote":"Provides COCO images used to construct the synthetic de-fencing datasets."}],"review_version":1}