Pith. sign in

REVIEW 3 major objections 5 minor 27 references

Dark Channel-Assisted Depth-from-Defocus from a Single Image

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 2506.06643 v2 pith:C36OCGNT submitted 2025-06-07 cs.CV

Moushumi Medhi , Rajiv Ranjan Sahay This is my paper
classification cs.CV
keywords depth-from-defocussingle-imagedepthestimationdarkchannelpriordefocusblurlocalvariationmapadversariallearningNYU-DepthV2
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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).

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

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.

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 (3)
  1. [Section III-A, Eqs. (4)-(5)] 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.
  2. [Section III-B, Fig. 6] 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.
  3. [Section III-B, Table I] 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.
minor comments (5)
  1. [Section I and Section III-C] 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.
  2. [Section II-A, Eq. (2)] 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.
  3. [Fig. 2 caption] 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.
  4. [Section III-A] 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.
  5. [Section II-B] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the dark channel and LDDCV features are computed directly from the defocused input image, depth is regressed against external NYU-v2 ground truth, and no equation in the paper reduces the predicted depth to its own inputs by construction.

full rationale

The derivation chain is a standard supervised pipeline with no step that is equivalent to its input by construction. The dark channel J_df (Eq. 1) and LDDCV map (Eq. 2) depend only on the defocused image I_df and its dark channel, never on ground-truth depth, so the central cue is an input feature rather than a fitted quantity. The depth output is trained with L_spafid = |d - d_gt|_1 against the NYU-v2 ground truth on the held-out eigen split, so the reported metrics are external rather than forced. Equation (5) is a forward thin-lens model used only to synthesize training blur and to condition the discriminator on the blur map r(d_gt)/r(d); it does not define the predicted depth. The adversarial term L_adv uses r(d) computed from the network's own estimate, but this is a consistency regularizer: the ablations (Table II, rows 1 and 2) show depth is predicted without the discriminator, so the output is not forced by the GAN loop. The reference list contains no self-citations, no uniqueness theorem is imported from the authors, and the pretrained backbone (ResNeXt [23]) and BSConv [24] are external components. The reviewer-flagged concerns, namely the possible dimensional inconsistency of Eq. 5 making the synthesized blur sub-pixel, the qualitative-only real-data evaluation (Fig. 6), and the paper's own conclusion that reliance on synthetic data may hinder generalization, are empirical correctness and external-validity issues rather than circularity; per the review rules they do not raise the circularity score. No specific reduction of any prediction to its inputs can be exhibited, so the honest finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method rests on the domain assumption that dark-channel local variation tracks defocus blur and that synthetic Gaussian blur approximates real lens blur; the hand-set parameters (T, loss weights, N, camera configuration) are not fitted to test data but shape the training distribution.

free parameters (4)
  • LDDCV validity threshold T = 0.05
    Hand-chosen threshold used to create the validity mask in MMSP-Net (Section II-B); no sensitivity analysis provided.
  • Loss weighting for frequency and adversarial terms = 0.1 each (Eq. 3)
    Fixed weights in L_total = L_spafid + 0.1 L_freq + 0.1 L_adv; chosen without ablation on these weights.
  • Number of nested feature modulation repetitions N = 2
    Selected by the authors as a memory/performance trade-off; no grid search shown.
  • Synthetic camera configuration (f, D_fp, F_n, p_x) = 9 mm, 0.7 m, 2, 7.5 µm
    Used to generate defocused training images via Eq. 5; these are arbitrary but plausible lens parameters and determine the entire training distribution.
assumptions (4)
  • domain assumption The dark channel's local variation decreases with increasing defocus blur.
    Motivated in Section II-A and Fig. 1 via a KDE plot of NYU-v2 depth values; treated as a general property of defocused images.
  • domain assumption Defocus blur is well approximated by a depth-dependent Gaussian PSF (Eq. 4) with radius r(d_gt) from the thin-lens formula (Eq. 5).
    Used to synthesize all training data from NYU-v2; real lens PSFs are more complex (e.g., spherical aberration, bokeh shape).
  • domain assumption Models trained on synthetic blur transfer to real defocused images.
    The paper evaluates on EBD real images without quantitative depth ground truth; the conclusion acknowledges synthetic training may limit generalization in dynamic or cluttered real environments.
  • domain assumption The adversarial discriminator's input (predicted depth, computed blur map, and defocused image) provides useful supervisory signal.
    Adversarial training is standard in image generation; here it is used empirically and the ablation shows a modest improvement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dark Channel-Assisted Depth-from-Defocus from a Single Image." pith.science (2026). https://pith.science/paper/C36OCGNT

@misc{pith2026250606643,
  author       = {Pith},
  title        = {Pith review of: Dark Channel-Assisted Depth-from-Defocus from a Single Image},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C36OCGNT}},
  note         = {Machine review of arXiv:2506.06643}
}
read the original abstract

We estimate scene depth from a single defocus-blurred image using the dark channel as a complementary cue, leveraging its ability to capture local statistics and scene structure. Traditional depth-from-defocus (DFD) methods use multiple images with varying apertures or focus. Single-image DFD is underexplored due to its inherent challenges. Few attempts have focused on depth-from-defocus (DFD) from a single defocused image because the problem is underconstrained. Our method uses the relationship between local defocus blur and contrast variations as depth cues to improve scene structure estimation. The pipeline is trained end-to-end with adversarial learning. Experiments on real data demonstrate that incorporating the dark channel prior into single-image DFD provides meaningful depth estimation, validating our approach.

Figures

Figures reproduced from arXiv: 2506.06643 by the authors.

Figure 1
Figure 1. Kernel Density Estimate Plot of Depth Values in the NYU-Depth [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the Dark Channel-Assisted DFD Framework. The diagram illustrates the key modules and workflow of the proposed dark channel-assisted [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Architecture of the Dark Channel-Infused Feature Boosting Unit. The schematic illustrates the structure of the DIFB unit, highlighting the integration [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Thin Lens Model of Defocus Blur. Defocus blur is modeled using [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]
Figure 5
Figure 5. Figure 5: Depth Estimation Results on synthesized defocused blur images from the NYU-v2 Dataset. (a) synthesized defocused blur images. (b)-(d) Estimated [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Depth estimation on high-resolution real defocused blur images from the EBD Dataset without fine-tuning. (a) Depth estimation results for two [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 26 canonical work pages

  1. [1]

    Extracting depth and radiance from a defocused video pair,

    X. Lin, J. Suo, and Q. Dai, “Extracting depth and radiance from a defocused video pair,”IEEE Trans. Circuits Syst. Video Technol., vol. 25, no. 4, pp. 557–569, 2014

    work page 2014

  2. [2]

    Discriminative filters for depth from defocus,

    F. Mannan and M. S. Langer, “Discriminative filters for depth from defocus,” inInt. Conf. 3D Vis. (3DV), 2016, pp. 592–600

    work page 2016

  3. [3]

    Depth map estimation using defocus and motion cues,

    H. Kumar, A. S. Yadav, S. Gupta, and K. Venkatesh, “Depth map estimation using defocus and motion cues,”IEEE Trans. Circuits Syst. Video Technol., vol. 29, no. 5, pp. 1365–1379, 2018

    work page 2018

  4. [4]

    Focus on defocus: bridging the synthetic to real domain gap for depth estimation,

    M. Maximov, K. Galim, and L. Leal-Taix ´e, “Focus on defocus: bridging the synthetic to real domain gap for depth estimation,” inIEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp. 1071–1080

    work page 2020

  5. [5]

    Self-supervised single-image depth estimation from focus and defocus clues,

    Y . Lu, G. Milliron, J. Slagter, and G. Lu, “Self-supervised single-image depth estimation from focus and defocus clues,”IEEE Robot. Autom. Lett. (RAL), vol. 6, no. 4, pp. 6281–6288, 2021

    work page 2021

  6. [6]

    Multi image depth from defocus network with boundary cue for dual aperture camera,

    G. Song, Y . Kim, K. Chun, and K. M. Lee, “Multi image depth from defocus network with boundary cue for dual aperture camera,” inIEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), 2020, pp. 2293– 2297

    work page 2020

  7. [7]

    Fully self-supervised depth estimation from defocus clue,

    H. Si, B. Zhao, D. Wang, Y . Gao, M. Chen, Z. Wang, and X. Li, “Fully self-supervised depth estimation from defocus clue,” inIEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 9140–9149

    work page 2023

  8. [8]

    Self-supervised spatially variant psf estimation for aberration-aware depth-from-defocus,

    Z. Wu, Y . Monno, and M. Okutomi, “Self-supervised spatially variant psf estimation for aberration-aware depth-from-defocus,” inIEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), 2024, pp. 2560–2564

    work page 2024

Show all 27 references
  1. [9]

    Deep depth from focal stack with defocus model for camera-setting invariance,

    Y . Fujimura, M. Iiyama, T. Funatomi, and Y . Mukaigawa, “Deep depth from focal stack with defocus model for camera-setting invariance,”Int. J. Comput. Vis., vol. 132, no. 6, pp. 1970–1985, 2024

    1970

  2. [10]

    Depth estimation and blur removal from a single out-of-focus image

    S. Anwar, Z. Hayder, and F. Porikli, “Depth estimation and blur removal from a single out-of-focus image.” inBrit. Mach. Vis. Conf. (BMVC), vol. 1, 2017, p. 2

    2017

  3. [11]

    Deep depth from defocus: how can defocus blur improve 3d estimation using dense neural networks?

    M. Carvalho, B. Le Saux, P. Trouv ´e-Peloux, A. Almansa, and F. Cham- pagnat, “Deep depth from defocus: how can defocus blur improve 3d estimation using dense neural networks?” inEur . Conf. Comput. Vis. (ECCV), 2018, pp. 0–0

    2018

  4. [12]

    Lens parameter estimation for realistic depth of field modeling,

    D. Pich ´e-Meunier, Y . Hold-Geoffroy, J. Zhang, and J.-F. Lalonde, “Lens parameter estimation for realistic depth of field modeling,” inIEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 499–508

    2023

  5. [13]

    Camera- independent single image depth estimation from defocus blur,

    L. Wijayasingha, H. Alemzadeh, and J. A. Stankovic, “Camera- independent single image depth estimation from defocus blur,” inIEEE Winter Conf. Appl. Comput. Vis. (WACV), 2024, pp. 3749–3758

    2024

  6. [14]

    Single image depth estimation trained via depth from defocus cues,

    S. Gur and L. Wolf, “Single image depth estimation trained via depth from defocus cues,” inIEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2019, pp. 7683–7692

    2019

  7. [15]

    Single image depth estimation using edge extraction network and dark channel prior,

    Y . Li, C. Jung, and J. Kim, “Single image depth estimation using edge extraction network and dark channel prior,”IEEE Access, vol. 9, pp. 112 454–112 465, 2021

    2021

  8. [16]

    Single image haze removal using dark channel prior,

    K. He, J. Sun, and X. Tang, “Single image haze removal using dark channel prior,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 33, no. 12, pp. 2341–2353, 2010

    2010

  9. [17]

    An enhanced window-variant dark channel prior for depth estimation using single foggy image,

    J. Chen and L.-P. Chau, “An enhanced window-variant dark channel prior for depth estimation using single foggy image,” inIEEE Int. Conf. Image Process. (ICIP), 2013, pp. 3508–3512

    2013

  10. [18]

    Underwater camera: Improving visual perception via adaptive dark pixel prior and color correction,

    J. Zhou, Q. Liu, Q. Jiang, W. Ren, K.-M. Lam, and W. Zhang, “Underwater camera: Improving visual perception via adaptive dark pixel prior and color correction,”Int. J. Comput. Vis., pp. 1–19, 2023

    2023

  11. [19]

    Image deblurring via extreme channels prior,

    Y . Yan, W. Ren, Y . Guo, R. Wang, and X. Cao, “Image deblurring via extreme channels prior,” inIEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2017, pp. 4003–4011

    2017

  12. [20]

    Deblurring images via dark channel prior,

    J. Pan, D. Sun, H. Pfister, and M.-H. Yang, “Deblurring images via dark channel prior,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 40, no. 10, pp. 2315–2328, 2017

    2017

  13. [21]

    Dark and bright channel prior embedded network for dynamic scene deblurring,

    J. Cai, W. Zuo, and L. Zhang, “Dark and bright channel prior embedded network for dynamic scene deblurring,”IEEE Trans. Image Process., vol. 29, pp. 6885–6897, 2020. 6

    2020

  14. [22]

    Indoor segmentation and support inference from rgbd images,

    N. Silberman, D. Hoiem, P. Kohli, and R. Fergus, “Indoor segmentation and support inference from rgbd images,” inEur . Conf. Comput. Vis. (ECCV), 2012, pp. 746–760

    2012

  15. [23]

    Towards robust monocular depth estimation: Mixing datasets for zero-shot cross- dataset transfer,

    R. Ranftl, K. Lasinger, D. Hafner, K. Schindler, and V . Koltun, “Towards robust monocular depth estimation: Mixing datasets for zero-shot cross- dataset transfer,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 3, pp. 1623–1637, 2020

    2020

  16. [24]

    Rethinking depthwise separable convolu- tions: How intra-kernel correlations lead to improved mobilenets,

    D. Haase and M. Amthor, “Rethinking depthwise separable convolu- tions: How intra-kernel correlations lead to improved mobilenets,” in IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp. 14 600–14 609

    2020

  17. [25]

    Depth and dof cues make a better defocus blur detector,

    Y . Jin, M. Qian, J. Xiong, N. Xue, and G.-S. Xia, “Depth and dof cues make a better defocus blur detector,” inIEEE Int. Conf. Multimedia Expo (ICME), 2023, pp. 882–887

    2023

  18. [26]

    P3depth: Monocular depth estimation with a piecewise planarity prior,

    V . Patil, C. Sakaridis, A. Liniger, and L. Van Gool, “P3depth: Monocular depth estimation with a piecewise planarity prior,” inIEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 1610–1621

    2022

  19. [27]

    Repurposing diffusion-based image generators for monoc- ular depth estimation,

    B. Ke, A. Obukhov, S. Huang, N. Metzger, R. C. Daudt, and K. Schindler, “Repurposing diffusion-based image generators for monoc- ular depth estimation,” inIEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 9492–9502

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.