{"id":"f4b1a9c6-2fc5-4497-bcb8-99ba3993f070","arxiv_id":"2506.06643","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Using the dark channel prior as a complementary input improves single-image depth-from-defocus, giving strong results on synthetic NYU-v2 blur and qualitative generalization to real blur.","lead":"A deep learning method estimates scene depth from a single blurred photo by adding the dark channel, the darkest pixels in small image patches, as an extra cue. A fixed-focus camera could then infer depth in one shot, with potential uses in robotics and mobile photography.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (5) is dimensionally inconsistent: with the stated camera parameters the blur radius is sub-pixel, so the synthetic training set may be effectively all-in-focus. The reported dark-channel gains would then not be evidence for depth-from-defocus.","rationale":"The reader identified synthetic-to-real generalization as the weakest assumption; I agree that is a concern, but the more immediate and internally checkable problem is whether the synthetic blur model is correct. If Eq. (5) has a typo and the code uses the standard thin-lens formula, the paper's ablation may be valid; if not, the quantitative core of the paper is undermined. The correct response is to require the authors to release the synthesis code and correct the equation, and to re-express the dark-channel gain in terms of defocus-specific cues. This does not contradict the reader's conditional verdict but sharpens its condition: the paper should not be accepted until this is resolved. I therefore recommend CONDITIONAL.","tokens_in":7833,"tokens_out":19449,"duration_ms":194889,"concrete_test":"Request the data-generation script or the exact blur radii used for the NYU-v2 subset; at a known depth (e.g., d_gt = 1 m) compute the effective Gaussian sigma of the PSF in the generated image. Compare with Eq. (5) (~5e-7 px) and with the corrected thin-lens formula (~1.2 px). If the code uses a corrected formula, the manuscript must fix Eq. (5) and state the intended units. If it uses Eq. (5) literally, retrain the full model and the DCC-ablated variant on images blurred with the corrected radius; if the AbsRel gap in Table II is not reproduced, the central claim that dark channel improves depth-from-defocus is unsupported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section III-A synthesizes defocus with Eq. (4) using a Gaussian PSF of standard deviation r, where r is defined by Eq. (5) as r = (1/sqrt(2)) * (p_x A / f) * (D_fp - f) * |d_gt - D_fp| / d_gt. With the paper's stated camera (f = 9 mm, F# = 2, D_fp = 0.7 m, p_x = 7.5 um), this evaluates to r ~ 5.5e-7 m at d_gt = 1 m. If r is treated as a pixel radius, as required by Eq. (4) where x,y are pixel coordinates, this is ~5e-7 pixels, i.e., essentially no blur; even if converted through p_x it is ~0.07 px. The standard thin-lens circle-of-confusion radius for the same settings is c/2 = f^2 |d - D_fp| / (2 F# (D_fp - f) d p_x) ~ 1.2 px. The equation as written has mismatched units (length rather than pixels) and is numerically about 10^4-10^6 times too small. If the experiments used Eq. (5) literally, the NYU-v2 images have negligible defocus; then Table II's improvement from the dark channel cue (0.118 -> 0.066 AbsRel) is attributable to ordinary image statistics, not defocus, and Table I does not exercise DFD at all. No code is released, so this cannot be checked from the paper. This is more basic than the synthetic-to-real generalization caveat the authors acknowledge: the synthetic data itself may not be defocused as described.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a supervised network for single-image depth-from-defocus. It augments a defocused RGB image with a dark channel and a Local Defocus and Dark Channel Variation (LDDCV) map, and trains the depth generator with frequency and adversarial losses. Experiments are performed on NYU-v2 images defocused with a Gaussian PSF using a thin-lens model, with a qualitative zero-shot test on the EBD dataset. The paper reports a state-of-the-art AbsRel of 0.042 on the synthetic NYU-v2 test set and an ablation showing large gains from the dark channel cue.","tokens_in":8209,"tokens_out":8242,"duration_ms":85375,"significance":"The contribution is potentially interesting: using the dark channel as a complementary defocus cue is a relatively novel idea in single-image DFD, and the ablation in Table II cleanly attributes much of the reported performance to that cue. However, the quantitative evidence rests on the synthetic defocus model in Eqs. (4)-(5), and that model is dimensionally inconsistent as printed. Because no code is released and the real-data evaluation is only qualitative, the central claim that the dark channel improves depth-from-defocus is not yet substantiated. The paper's strengths are the clear architecture description and the controlled ablation; its weakness is the unresolved validity of the synthetic data-generation step.","major_comments":[{"comment":"Equation (5) is dimensionally inconsistent. With the stated camera parameters (f = 9 mm, F# = 2, D_fp = 0.7 m, p_x = 7.5 um), the expression evaluates to approximately 5.5e-7 m at d_gt = 1 m. Since Eq. (4) is applied in pixel coordinates, r must be expressed in pixels; if one converts the raw value through p_x, r is about 0.073 px, and if one uses the raw value directly as a pixel radius, r is about 5.5e-7 px. Either way, the synthesized NYU-v2 images contain essentially no defocus, so Tables I and II would not exercise depth-from-defocus at all. This is a load-bearing issue for the paper's central claim. Please correct the formula, clearly state the conversion to pixel units, and supply code or a quantitative verification (for example, a histogram of r over the training set) demonstrating that the synthetic defocus is non-negligible.","section":"Section III-A, Eqs. (4)-(5)"},{"comment":"The real-data evaluation is qualitative only. Figure 6 shows two examples without any quantitative metrics, and the EBD dataset has no ground-truth depth, but the abstract concludes that 'experiments on real data demonstrate meaningful depth estimation,' which is stronger than the evidence presented. If real-data validation is claimed, the authors should provide a quantitative protocol, such as ordinal accuracy, comparison with predicted defocus magnitude, or a human evaluation, or they should explicitly temper the real-data claim.","section":"Section III-B, Fig. 6"},{"comment":"Several comparisons are not on equal footing and should be disclosed more prominently. The BC-DAC numbers were provided by the original authors rather than rerun, Camind was evaluated on a different depth range than in its original paper, and P3Depth and Marigold are all-in-focus methods. These caveats do not invalidate the conclusion that the proposed method is competitive, but the table and text should clearly distinguish rerun results from author-provided results and state the evaluation range for each baseline.","section":"Section III-B, Table I"}],"minor_comments":[{"comment":"There are typographical and formatting artifacts, including 'TThese methods' in Section I and 'ding174'/'ding72' in Section III-C, which appear to be broken cross-references and should be fixed.","section":"Section I and Section III-C"},{"comment":"Equation (2) is not cleanly defined: the notation '∥p=i-1, i, i+1, q=...' is nonstandard and should be rewritten with an explicit maximum over the 3x3 neighborhood.","section":"Section II-A, Eq. (2)"},{"comment":"The caption for Fig. 2 refers to 'image enhancement,' but the task is depth estimation; the caption should be corrected to describe the depth-from-defocus pipeline.","section":"Fig. 2 caption"},{"comment":"The relationship between the variable r in Eq. (5) and the standard deviation of the Gaussian PSF in Eq. (4) should be stated explicitly, including the pixel-unit conversion, even after the formula is corrected.","section":"Section III-A"},{"comment":"The threshold T = 0.05 for the validity mask is introduced without motivation or sensitivity analysis; a sentence explaining its choice or a reference to an ablation would be helpful.","section":"Section II-B"}],"recommendation":"major_revision","confidential_remarks":"The dimensional inconsistency in Eq. (5) is the central technical concern. I recommend major_revision rather than reject because a typographical error in the formula is plausible and correctable. However, without code or a clear statement of the actual blur radii used, the current manuscript's central experimental claim is unsupported. I would ask the editor to require the authors to provide code or a detailed verification of the defocus synthesis as a condition of revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper is worth a referee's time but not yet worth believing. The core idea—using the dark channel's local variation as an extra cue for single-image depth-from-defocus—is genuinely new, and the ablation (Table II) is set up to isolate exactly that contribution. If the synthetic benchmark held up, AbsRel 0.042 with δ1 0.975 would be a strong result for a niche subfield. The authors also compare against a healthy set of baselines and are transparent about which numbers came from the original authors (BC-DAC).\n\nThe problem is in Section III-A. Equation (5) as printed is dimensionally inconsistent. Plugging in their own camera parameters (f=9mm, F#=2, D_fp=0.7m, p_x=7.5µm) gives a blur radius r of about 5.5e-7 m; even converted to pixels, that is roughly 0.07 px. A Gaussian with that standard deviation is invisible. The correct thin-lens CoC radius for the same settings is of order 1 px. So either the equation is a typo and they used the right formula in code, or the NYU-v2 training set is effectively all-in-focus and the reported gains have nothing to do with defocus. With no code released and no visible kernel statistics, the reader cannot tell which. That is a load-bearing ambiguity, not a cosmetic one.\n\nThe rest of the evaluation is consistent with the reader's report: real-data results are qualitative only, no error bars, and one baseline is represented by numbers supplied by its own authors. Those are familiar soft spots. The dimensional issue is the one that needs to be resolved first.\n\nIf the authors can supply the corrected equation, code, and a quick empirical check that their synthetic blur is actually observable, the paper becomes citable. As it stands, I would not cite the quantitative results. I would still send it to reviewers—the idea is worth a careful look and the flaw is very likely fixable—but the referee should be asked to verify Eq (5) before anything else.\n\nBring it to the reading group if you want to talk about how a single bad equation can sink an otherwise plausible empirical story. Otherwise, wait for the revised version.","headline":"Interesting idea and clean ablation, but Eq (5) may mean the synthetic benchmark never actually exercised defocus; needs correction and code before the DFD claim is credible.","tokens_in":8760,"tokens_out":6443,"would_cite":false,"duration_ms":59673,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Single-image depth-from-defocus succeeds when the dark channel's local intensity variation is added as a complementary cue, and the resulting network beats single-defocus, focal-stack, and all-in-focus baselines on NYU-v2.","keywords":["depth-from-defocus","single-image depth estimation","dark channel prior","defocus blur","local variation map","adversarial learning","NYU-Depth V2"],"falsifier":"A concrete experiment would be capturing a set of real scenes with a shallow-DoF camera and a depth sensor (or structured-light ground truth), then comparing the proposed model's predicted depth against an all-in-focus monocular baseline on those images; if the model does not beat that baseline, the synthetic-to-real transfer claim fails.","tokens_in":7618,"feed_emoji":"📷","tokens_out":5924,"duration_ms":54056,"temperature":0.7,"pith_summary":"The paper tries to show that a single defocused photograph carries enough information for depth estimation, provided the right cue is made explicit. That cue is the dark channel: the per-pixel minimum over color channels in a local window, which retains edges and scene structure where blur is small and is smoothed where blur is large. The authors define a two-channel Local Defocus and Dark Channel Variation (LDDCV) map from the defocused image and its dark channel, and feed it into an encoder-decoder network trained end-to-end with an adversarial loss. On synthetic defocus generated from NYU-v2 images, the full model reaches an absolute relative error of 0.042 and a $\\delta_1$ accuracy of 0.975, and the ablation attributes most of the gain to the dark channel cue (AbsRel improving from 0.118 to 0.066 when it is added).","feed_headline":"Dark channel cue turns one blurry photo into a depth map","feed_subtitle":"Adding the dark channel's local variation beats single-defocus, focal-stack, and all-in-focus baselines on NYU-v2 depth errors.","key_machinery":"The load-bearing object is the Local Defocus and Dark Channel Variation (LDDCV) map, defined at each pixel as the pair of maximum absolute intensity differences between the pixel and its 3x3 neighbours, computed once on the defocused image and once on its dark channel. Because defocus blur homogenizes local regions, these maxima fall where blur is strong and stay high where the scene is near the focal plane, giving a depth-dependent signal. The network couples this map to a pretrained ResNeXt encoder through an LDDCV embedding net and a mask-mediated sparse pooling net, while a dark channel embedding net and nested feature-modulation blocks inject the dark channel's structural information into the decoder. A discriminator that sees predicted or ground-truth depth together with the corresponding defocus blur map supplies adversarial supervision during training.","core_discovery":"The central discovery claimed is that defocus blur, which suppresses local intensity variation, leaves a readable trace in the dark channel, and that trace can be turned into depth. In the paper's formulation, blur homogenizes local regions, so high-defocus areas have low local variation and in-focus areas retain high variation in both the image and its dark channel. Concatenating these two local-variation maps into the LDDCV map, and modulating deep features with an embedding of the dark channel, lets a single network regress a depth map from one blurred RGB image. The authors report that this outperforms prior single-defocus methods, a focal-stack method, and two all-in-focus baselines on most NYU-v2 metrics, and that the dark channel is the dominant component in the ablation.","pith_inferences":["An extension the paper leaves implicit: because the LDDCV map is cheap to compute, the dark channel cue could be packaged as a lightweight auxiliary input for mobile or embedded depth-from-defocus systems without a ResNeXt-scale encoder.","A testable extension would be training the same architecture on real defocused images with depth maps produced by a depth sensor, to see whether the dark channel gain persists when the Gaussian PSF assumption is removed.","Since the dark channel responds to any smoothing that destroys local minima structure, similar local-variation maps may help for motion-blurred or low-light images, not only defocus."],"forward_implications":["A fixed-focus, wide-aperture camera can infer depth passively from a single frame, with no autofocus sweep, multi-image capture, or active illumination.","The dark channel, previously a tool for haze removal and deblurring, becomes a practical monocular depth cue for defocus blur.","Adversarial supervision conditioned on the defocus blur map improves depth accuracy over the same architecture trained without it (AbsRel from 0.077 to 0.042 in the ablation).","Performance on real high-resolution defocused images without fine-tuning indicates the learned cue transfers across blur magnitudes, at least qualitatively."],"supporting_citations":[{"why":"Supplies the 795/654 train/test RGB-D pairs whose all-in-focus images are blurred to create the synthetic defocus training data.","marker":"[22]"},{"why":"Single-defocus baseline trained and tested under the same split; its depth predictions are compared against in Table I and Fig. 5.","marker":"[11]"},{"why":"Camera-independent single-image DFD baseline that also uses defocus blur supervision; comparison in Table I.","marker":"[13]"},{"why":"Dual-aperture multi-image depth-from-defocus baseline whose reported results are used as a strong comparison in Table I and Fig. 5.","marker":"[6]"},{"why":"Introduces the dark channel prior, the definition and intuition the paper adapts from haze removal to defocus depth.","marker":"[16]"},{"why":"Establishes that dark channel statistics respond to blur, supporting the claim that dark channel variation is a defocus cue.","marker":"[20]"},{"why":"Provides the real high-resolution defocused images used to test zero-shot generalization without fine-tuning.","marker":"[25]"},{"why":"Provides the pretrained ResNeXt101 encoder backbone that extracts multi-scale features from the defocused image.","marker":"[23]"},{"why":"Defines the blueprint separable convolutions used throughout the generator to reduce parameters by about 49 percent.","marker":"[24]"}],"fun_headline_variants":["Dark channel reveals depth from single-image defocus","Blur's dark channel maps depth from a single shot","One defocused image's dark channel yields depth","Depth from single blur via dark channel clue","Dark channel in defocus blur recovers depth"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that the synthetic Gaussian defocus blur used for training behaves enough like real lens blur that learned cues transfer to real photographs; if the blur model is too idealized, the method's edge over baselines may not hold on real optics.","fun_headline_variants_meta":{"raw":{"variants":["Dark channel reveals depth from single-image defocus","Blur's dark channel maps depth from a single shot","One defocused image's dark channel yields depth","Depth from single blur via dark channel clue","Dark channel in defocus blur recovers depth"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000652,"raw_usage":{"total_tokens":2934,"prompt_tokens":837,"completion_tokens":2097,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":2025}},"tokens_in":453,"tokens_out":2097,"duration_ms":13871,"temperature":1.0,"reasoning_tokens":2025,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:51:56.943044+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete experiment would be capturing a set of real scenes with a shallow-DoF camera and a depth sensor (or structured-light ground truth), then comparing the proposed model's predicted depth against an all-in-focus monocular baseline on those images; if the model does not beat that baseline, the synthetic-to-real transfer claim fails.","supporting_citations":[{"cited_title":"Indoor segmentation and support inference from rgbd images,","cited_arxiv_id":null,"evidence_quote":"Supplies the 795/654 train/test RGB-D pairs whose all-in-focus images are blurred to create the synthetic defocus training data."},{"cited_title":"Deep depth from defocus: how can defocus blur improve 3d estimation using dense neural networks?","cited_arxiv_id":null,"evidence_quote":"Single-defocus baseline trained and tested under the same split; its depth predictions are compared against in Table I and Fig. 5."},{"cited_title":"Camera- independent single image depth estimation from defocus blur,","cited_arxiv_id":null,"evidence_quote":"Camera-independent single-image DFD baseline that also uses defocus blur supervision; comparison in Table I."},{"cited_title":"Multi image depth from defocus network with boundary cue for dual aperture camera,","cited_arxiv_id":null,"evidence_quote":"Dual-aperture multi-image depth-from-defocus baseline whose reported results are used as a strong comparison in Table I and Fig. 5."},{"cited_title":"Deblurring images via dark channel prior,","cited_arxiv_id":null,"evidence_quote":"Establishes that dark channel statistics respond to blur, supporting the claim that dark channel variation is a defocus cue."},{"cited_title":"Depth and dof cues make a better defocus blur detector,","cited_arxiv_id":null,"evidence_quote":"Provides the real high-resolution defocused images used to test zero-shot generalization without fine-tuning."},{"cited_title":"Towards robust monocular depth estimation: Mixing datasets for zero-shot cross- dataset transfer,","cited_arxiv_id":null,"evidence_quote":"Provides the pretrained ResNeXt101 encoder backbone that extracts multi-scale features from the defocused image."},{"cited_title":"Rethinking depthwise separable convolu- tions: How intra-kernel correlations lead to improved mobilenets,","cited_arxiv_id":null,"evidence_quote":"Defines the blueprint separable convolutions used throughout the generator to reduce parameters by about 49 percent."}],"review_version":1}