acceptodds
Under review as a conference paper at ICLR 2027

From Wall Glows to Recoverable Screens: Hierarchical Recoverability-Guided Accretive Representation Learning for Indirect Screen Reconstruction

Abstract

We study whether a hidden display can be recovered from an ordinary camera image of the diffuse glow that it casts onto a visible wall. Spatial propagation and wall reflection mix screen regions and attenuate color, layout, structure, and detail by different amounts at different locations, so each wall glow retains a different set of usable cues. A uniform image-to-image mapping leaves the recoverability and reconstruction strength of these traces unresolved. We introduce Observability-Guided Accretive Representation Learning (OARL) to make these decisions from the observation. Calibrated Hierarchical Recoverability Representation Learning (CHRRL) constructs a shared evidence field, organizes it into color, layout, structure, and detail, and predicts a spatial gate for each type. A paired-versus-nonmatching screen objective calibrates the global evidence representation during training. Observability-Stratified Accretive Reconstruction (OSAR) then follows a fixed color-to-layout-to-structure-to-detail dependency backbone; the gates continuously modulate each stage's residual magnitude at every location, and later stages inherit earlier states. At inference, the complete path uses the current wall-glow image. OARL achieves the best reported values in all 15 category–metric comparisons across five screen-content categories. Component and controlled representation comparisons support hierarchical recoverability estimation and inherited gated reconstruction, while the seven-position camera-translation series shows stable recovery as the observed glow changes. Together, these results provide initial evidence that an ordinary camera viewing a wall glow can recover the principal information of a hidden screen.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Indirect Screen Reconstruction, Recoverability Estimation, Representation Learning, Progressive Reconstruction, State Inheritance

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_5CYzOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper studies whether the content of a hidden screen can be reconstructed from a camera image of the diffuse light reflected onto a nearby wall. It proposes OARL, a framework that estimates the recoverability of different types of screen information (color, layout, structure, and detail) and uses these estimates to guide a progressive reconstruction process. The evaluation includes five categories of screen content, comparisons with alternative approaches, ablation studies, and experiments under different imaging conditions. The reported results are encouraging, although I have some reservations about how clearly the experiments establish the benefits and generality of the proposed design.

Critical strengths

[+] The paper explores an unconventional reconstruction setting and provides a physical acquisition setup with paired observations and screen images. This gives the work a clear experimental foundation.

[+] The authors include multiple content categories, ablations, and robustness experiments. These experiments provide a useful initial assessment of the proposed approach.

Critical weaknesses

[-] The paper introduces several components and a fixed reconstruction order, but the current experiments do not make it entirely clear how much each design choice contributes beyond a conventional progressive reconstruction approach.

[-] The experiments mainly support reconstruction under the tested acquisition conditions. More direct evaluation of text, icons, and other fine-grained content would help establish what the method can reliably recover.

Questions for authors
  1. Could the authors clarify how the training, validation, and test sets are separated? In particular, are the screen contents and acquisition sessions independent across these sets?

  2. Could the authors provide a more controlled comparison with a conventional reconstruction model using the same backbone and training setup, to better isolate the contribution of the proposed gates and progressive reconstruction?

  3. How well does the method perform when the camera, wall, or display differs from the acquisition setup used for training?

  4. Do the authors have an evaluation of fine-grained content accuracy, such as text or icon recognition, beyond the reported image similarity metrics?

4 comments

We thank Reviewer 5CYz for the updated review. The initial review was posted on October 19, 2026. The review page shows that it was updated on November 14, 2026. The initial review requested additional evidence on held-out screen identities, unseen acquisition conditions, and content-level fidelity. We completed the corresponding evaluations using the available data and an additional second-room set; the results are summarized in Part 1. The revised review raises four related but distinct questions, which we address in Part 2.

Part 1. Response to the initial review. The three evaluations requested in the initial review are complete, and their results support the paper. We added evaluations of screen and template separation, transfer to a second room, and content-level reconstruction. Table S6 reports zero test pairs sharing a source screen or template with the training portion. Table S7 reports results on a template-grouped Screen split, where OARL reaches 25.472 dB PSNR and exceeds the strongest retrieval baseline by 6.059 dB. Table S14 evaluates the Table 1 weights on 200 Screen pairs from a second room with a different wall, camera body, and camera distance, and OARL leads on all five reported measures, with 23.107 dB against 21.142 dB for Restormer and 17.859 dB for retrieval. Table S12 reports layout IoU and text-region recall, and OARL leads on both, with 0.77 and 44% on Screen against 0.66 and 22% for the strongest restoration baseline.

Part 2. Response to the revised review. Each of the four questions of the revised review has a direct answer in our data.

  1. Data partition and independence. Training and test are cleanly separated at the level of source screens and templates, and the second-room set adds a separate capture session. Table S6 reports the train and test counts and shows that no test pair shares a source screen or template with the training portion. Table S7 adds a Screen split in which each template group is assigned wholly to training or test, with OARL, restoration, and retrieval comparisons. Section E.9 describes the screen-display and image-capture procedure. We showed each screen for 3 s by script and photographed it once from a tripod with fixed exposure, ISO, white balance and focus, so every pair links one photograph to one screen. The second-room set of Table S14 comes from a capture session separate from all training data. These results establish source-screen and template separation.

  2. Contributions of the gates and progressive reconstruction. The gain comes from the gates and the progressive reconstruction, and it remains when the backbone is fixed and when the capacity is matched. Table 2 evaluates component removals and replacements. On Screen the full model reaches 25.56 dB, against 23.90 dB without the gates and 22.87 dB without inheritance. Tables 3 and 4 compare representation and learning configurations under a shared U-Net recovery architecture. Figure 6 presents inheritance and reconstruction-order comparisons, and Table S9 adds a capacity-matched Restormer comparison, in which Restormer reaches 22.705 dB against 25.472 dB for OARL. Together, these results support contributions from the proposed components and show that the gains are not explained by model size alone.

  3. Changes in acquisition conditions. The Table 1 weights transfer to a second room and stay ahead under camera-position changes. Table S14 evaluates the Table 1 weights on 200 Screen pairs from a second room with a different wall, camera body, and camera distance. OARL stays 1.97 dB above Restormer and 5.25 dB above retrieval there, and its layout IoU is 0.73 against 0.77 in the training room. Figures 7 and 8 show quantitative and qualitative results under camera-position changes, where we moved the camera through seven positions and OARL leads on all three metrics at each one. This shows transfer to a second room and stability across camera positions.

  4. Content-level accuracy. The screen content is recovered and measured directly, and OARL leads on every measure. Table S12 reports layout IoU and text-region recall for Screen, Browser, and OSLock, and Table S14 reports the corresponding measures for the second-room set. On Screen OARL reaches 0.77 and 44% against 0.59 and 16% for retrieval, and in the second room 0.73 and 37% against 0.48 and 6%. Exact character transcription and icon identification from a diffuse wall have not been reached by any published method. Backes et al. (IEEE S&P 2009) studied this same channel and resolved a single 10 cm letter on a 3×3 grid at 25 cm, and the readable text in the literature comes from mirror-like reflections or from the raster timing of a CRT. Our first response gives this literature in full under (c).

The method is motivated by the structure of the inverse problem. A wall observation spatially mixes contributions from screen regions and attenuates different spatial frequencies and content types unequally. A uniform reconstruction rule does not explicitly represent this variation. CHRRL estimates spatially and semantically specific support from the current observation, while OSAR uses the resulting gates to regulate residual updates and carries earlier color and layout states into later structure and detail stages. Better-supported cues thereby guide the updates that depend on weaker evidence.

Several completed results support this design. In Table S10, the complements of all four gates correlate with reconstruction errors more strongly than the single-attention control across the five categories. Table S11 shows that gate responses decrease under stronger noise and motion blur, with the detail gate showing the largest decrease. Table S7 shows that OARL retains an advantage over the evaluated retrieval baselines on a template-grouped split, Table S9 shows an advantage over a capacity-matched Restormer, and Table S14 shows transfer to an additional room using the Table 1 weights. For content-level evaluation, Table S12 measures layout IoU and text-region recall.

We appreciate that Reviewer 5CYz recognizes the physical acquisition setup and paired wall and screen data as providing a clear experimental basis, and describes the multi-category, ablation, and robustness experiments as a useful initial evaluation. The evidence above supports recovery of the major colors, layouts, and structures of the screen.

We would value the view of the reviewer on whether the results in Part 1 address the concerns raised in the initial review and whether Part 2 addresses the questions of the revised review.

Thank you for recognizing the task as a meaningful challenge and the division between CHRRL and OSAR as understandable.

Your review asks for three evaluations, namely held-out screen identities, unseen acquisition configurations, and the fidelity of semantically important details. The revised paper contains all three. Every number is given in this reply, and the same tables are in Appendix E of the revised PDF.

(a) Held-out screen identities. Every test screen is new to the model. The number of test pairs that share a source screen or a template with training is 0 in all five categories (Table S6). We also built a second split of Screen that places each template group entirely in training or entirely in test, retrained OARL and its strongest baselines on it, and added retrieval. Retrieval returns the training screen whose wall observation is nearest to the test observation. Its output is a real, sharp, plausible screen, and it is the most that training content alone can supply, so it is the reference for a model that fills in content from learned priors.

Screen, template-grouped splitPSNRSSIMLPIPS
OARL25.4720.8490.222
Restormer22.3660.7680.348
Restormer, matched22.7050.7730.329
Best retrieval19.4130.6870.481

OARL reproduces Table 1 within 0.086 dB and stays 6.06 dB above the best retrieval. This 6 dB is what OARL takes from the current wall observation beyond the training content. Over three seeds OARL reaches 25.463±0.073 dB against 22.350±0.096 dB for Restormer.

(b) An unseen acquisition configuration. OARL keeps its lead in a room it has never seen. We applied the Table 1 weights directly to 200 Screen pairs from a second room. The room, the wall, the camera body and the camera distance are all different, the screens lie outside the training set, and the session is separate from all training data.

Second room, Table 1 weightsPSNRSSIMLPIPSLayout IoU
OARL23.1070.7970.3160.73
Restormer21.1420.7190.4280.64
Retrieval of the nearest training screen17.8590.6270.5940.48

OARL leads on every measure, by 1.97 dB over Restormer and 5.25 dB over retrieval. Its layout IoU is 0.73 in the second room and 0.77 in the training room. The same weights also carry across content. The Screen model applied directly to Browser, whose pages are all new to it, reaches 16.967 dB against 13.774 dB for Restormer.

(c) Fidelity of semantically important content. OARL recovers the layout and the text regions of the screen better than every baseline. The paper claims recovery of screen colors, regional layouts and principal structures (abstract), and we now score this content directly. Layout IoU scores the regional layout of each output against the true screen. Text-region recall scores whether text sits where the true screen has text.

Layout IoU / text-region recallScreenBrowserOSLockScreen, second room
OARL0.77 / 44%0.72 / 25%0.60 / 20%0.73 / 37%
Strongest restoration baseline0.66 / 22%0.65 / 17%0.57 / 18%0.64 / 18%
Retrieval0.59 / 16%0.55 / 10%0.48 / 8%0.48 / 6%

OARL leads on both measures in all four settings. On Screen it places twice as many text regions as the strongest restoration network, and in the second room six times as many as retrieval.

Exact character recognition and icon identity are beyond what the diffuse glow of a display carries for every published method. Backes et al. (IEEE S&P 2009) studied this same channel with a monitor facing a white wall 25 cm away. With physical deconvolution they resolved a single 10 cm letter on a 3×3 grid and concluded that diffuse reflections reveal no more than basic shapes. The readable text in this literature comes from other signals, namely mirror-like reflections in teapots, glasses and the eye (Backes et al., IEEE S&P 2008) and the raster timing of a CRT read with a fast photosensor (Kuhn, IEEE S&P 2002). A photograph of a matte wall lit by a flat panel contains neither. Passive imaging from one ordinary photograph relies on the penumbra of an occluder (Saunders et al., Nature 2019), and a 2026 study by Raji and Murray-Bruce (arXiv 2601.12257) describes such methods as still confined to one-dimensional or low-resolution two-dimensional color images. The learned methods NLOS-OT and NLOS-LTM reconstruct single digits, faces and small natural images, and NLOS-LTM reports PSNR and SSIM only. None of these works reports character or icon recognition from a diffuse wall. Our evaluation adds two content-level measures to the usual image metrics, layout IoU for where the regions of the screen are and text-region recall for where its text is, and OARL leads on both.

The five questions are answered in order.

  1. No test screen and no test template appears in training. The split is described in (a). Each category is divided into a training portion and test pairs, 3,040/760 for Screen, 1,000/250 for Figure, 1,000/250 for Game, 1,120/279 for OSLock and 3,759/940 for Browser. Source screenshots and templates are disjoint between the two. A validation subset held out from the training portion selects checkpoints, and the test pairs are used for the reported numbers.

  2. A recoverability gate makes two predictions, and both are confirmed. Where the gate is low, the reconstruction error is high. The direct control for an ordinary learned attention map is a single attention map trained in the same network. Its correlation with the high-frequency reconstruction error is 0.09 to 0.15, the detail gate reaches 0.42 to 0.51, and every one of the four gates exceeds the attention map on every error measure in all five categories (Table S10). When information is removed from the observation, the gate falls. We degraded the observation in controlled steps of brightness, noise and blur and recorded each gate (Table S11).

ObservationColorLayoutStructureDetail
Nominal0.8670.7920.6890.601
Noise, SNR 10 dB0.7810.6400.4320.231
Motion blur, 15 px0.8030.6640.4710.259

Under both noise and blur the gates fall in the order color, layout, structure, detail, by 10%, 19%, 37% and 62% at the strongest noise. The wall loses detail first and color last, and this measured order is the order in which OSAR writes.

  1. The gates and the inherited stages carry the gain. Table 2 makes the replacement you describe inside OARL, on all five categories. On Screen, OARL with the gates removed gives 23.90 dB and with inheritance removed gives 22.87 dB, and the conventional replacements of either module (a convolutional gate, a ResNet refinement head and four others) give between 16.58 and 20.26 dB. Full OARL gives 25.56 dB. Against a separate architecture, a Restormer matched in size, input, losses and training budget reaches 22.705 dB on the template-grouped split, and OARL stays 2.767 dB above it. OARL has 17.1M parameters and the standard Restormer of Table 1 has 26.1M.

  2. The model works in an unseen room with its weights unchanged. Test (b) is this evaluation. One set of weights, trained in the first room, is evaluated as it is in the second room, and OARL leads on all five measures of Table S14.

  3. OARL leads every baseline on the recovered content. Test (c) is this evaluation, on Screen, Browser, OSLock and the second room (Tables S12 and S14).

We will release the dataset, splits, weights and code. We hope the three evaluations you asked for resolve your concerns, and we would be grateful if you would consider raising your score.

Reviewer_YbKzOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper tackles the problem of reconstructing a computer screen that is hidden from view by analyzing the reflection of the screen off of a white diffuse wall. This is a challenging problem that requires deconstructing the conflated signals and reconstructing the original screen.

The authors propose a two-stage method of reconstructing the screen. The first stage analyzes the input image of the wall reflection and generates an estimate of where information about the screen is likely to be found. The authors break this information into four parts: color, layout, structure, and detail. The first stage outputs an attention map for each of the four information categories.

The second stage builds up a reconstruction of the scene step by step, going through the evidence for each of the four categories serially. Each step adds information to the reconstruction, until the reconstruction is complete. The process is not iterative, but each step adjusts and hopefully improves the reconstruction.

The authors compare their method against a large number of other networks trained for image restoration or enhancement. Their design is able to set a new quantitative standard for the task, and the qualitative results are impressive.

Critical strengths

The two stage method appears to be a useful design, given the ablation study. Stage 2, the OSAR module, is similar in nature to prior work on intrinsic imaging, "Colorful Diffuse Intrinsic Image Decomposition in the Wild" by Careaga and Aksoy, SIGGRAPH 2024. They reconstruct the intrinsic images using a multi-stage process (see their figure 3). It is not exactly the same design as this paper, but the concept is similar in that they build up the intrinsic image outputs serially by estimating different attributes of the output at each stage.

The stage 1 module, CHRRL, which estimates where evidence is likely to occur for different properties in the input images, is also an interesting design. This type of module has potential uses in other image enhancement or analysis tasks.

The quantitative and qualitative results are also a strength of the paper. The qualitative results are particularly impressive. The quantitative results are good, but you don't understand how good without seeing actual reconstructions.

Critical weaknesses

There are no obvious critical weaknesses in the paper. However, there are numerous small issues with the clarity of the writing, the figures, and the equations. A frustrating part of reading the paper is that the most interesting aspects of it are difficult to understand from the figures, equations, and descriptive text. I've included a number of specific items below that should be corrected. In particular, figure 2 contains duplicated pathways that look different, since the CHRRL and OSAR modules are both shown in detail across the top panel and the bottom panel of figure 2. However, those details are different in the two panels, which creates a lot of confusion.

Below are my specific comments and thoughts while reading the paper.

Good abstract, clearly written

Line 075: colons and semi-colons are showing up with extremely high frequency in papers I'm reviewing, and they usually make the sentences more difficult to parse mentally. Conjunctions, commas, or dividing the sentence into two full sentences usually read much more cleanly. One other example where a colon is not necessary is line 085 (first contribution sentence). Just make that two sentences so that the first clause is a clean statement.

Line 110: Invalid citation at the start of the sentence.

Line 138: 'Figure2' needs a space.

Line 160: '...updates through semantic path.' -> 'semantic paths' or 'a semantic path'?

line 158: 'f retains local/contextural support', what is f? It's not used in (1) and not defined in the section. It is shown in figure 2, but it still needs to be defined in the text if you're going to use it.

Line 161: 'Fixed order; evidence governs update strength' means what? Please ditch the semi-colon and make this a meaningful sentence. What has a fixed order? The paragraph already starts with 'Evidence guides reconstruction'. What does the second clause add? What does it mean?

Figure 2: the information in the top section is mostly duplicated in the bottom section. I would expect the blowup of the top section to contain more information, but both the top and bottom sections contain roughly the same scale of information, just with different arrows and symbols, which makes it very confusing. A better structure might be to have a more abstract figure representing the information degradation abstraction box and then have simple boxes for the CHRRL and OSAR components. Then provide CHRRL and OSAR each a figure of their own so that the connections and information flow are shown only once and not as tiny as the current figure.

Equation 2: you're working in a discrete image domain, so integrals aren't really appropriate. This is true for all of the equations. The data is not continuous.

Equation 3: what is A_vt^(f)?

Equation 13: the text describes d, l, and Delta, but there is no l symbol in equation 13. Should l be L_1? Is L the final loss?

Line 198: another gratuitous semi-colon

Line 301: 'Table 1 summarizes main results.' -> 'summarizes the main results' ?

Table 1: what is the meaning of the blue lines? It looks like these are the models with at least one 2nd place metric. What about just highlighting the specific metric? It's not easy to spot blue text within a blue line to figure out which metric is relevant.

Figure 3: the images are so small. What about reducing the figure to just the top 5-7 performing methods and making the images bigger?

Figure 6: again, how about including just the top 5-7 methods?

Questions for authors

One issue with the staged reconstruction path is that an error early in the path can't be corrected later. Is there a mechanism in the OSAR to minimize that issue? Could a later stage (e.g. structure) correct errors in an earlier stage (e.g. color or layout)?

How would you make the case that something like the CHRRL module could be generalized to other tasks?

What does a failure case look like?

How specific is the model to the particular imaging configuration in the training data? Would it work in a different room with a slightly different imaging configuration or screen?

1 comment

Thank you for the careful reading and the detailed line notes.

  1. A later stage can correct an earlier one. Every stage reads the shared field f together with all earlier states (Eq. 8) and adds its own residual to the same reconstruction (Eq. 10), so the structure stage adjusts what the color and layout stages wrote at any pixel. Each residual is scaled by its gate, so every stage writes in proportion to its evidence.

  2. CHRRL needs two inputs that most restoration tasks already have. The module reads local statistics and gradients of an observation and a preliminary estimate (Eqs. 1 to 5). Dehazing and low-light enhancement offer the same two inputs and the same uneven evidence across the image. On our data each gate follows the reconstruction error more closely than a single attention map in all five categories (Appendix E.5).

  3. Appendix E.8 of the revised PDF shows three cases from Browser and OSLock, each with the wall observation, the true screen and the reconstruction side by side.

  4. The model works in a second room with the Table 1 weights. The second-room set has 200 Screen pairs with a different wall, camera body and camera distance, and its screens lie outside the training set (Appendix E.11). PSNR/SSIM/LPIPS are 23.107/0.797/0.316 for OARL against 21.142/0.719/0.428 for Restormer.

  5. Figure 2 is now a single overview that runs from the wall observation to the preliminary estimate, to CHRRL and its four gates, to OSAR and the output. Figure 3 in Section 3.2 shows CHRRL alone, from y and z through the local statistics of Eq. 2 and the evidence field of Eq. 3 to the four gates of Eq. 5. Figure 4 in Section 3.3 shows OSAR alone, from z through the four stages of Eqs. 8 to 10 to the bounded output of Eq. 12. The paragraph at lines 158 to 161 now reads as follows.

Evidence guides reconstruction throughout the model. The estimate z anchors the gates to screen coordinates, the evidence field f of Equation (3) retains local and contextual support, CHRRL produces four maps m, and OSAR scales the stage residuals at the same locations. Reliable traces initialize the states and give context to weaker structure and detail. Recoverability thus enters every update through the semantic path, which limits the amplification of weak or mixed evidence. The order of the four stages is fixed, and the gates set the strength of each update.

f is the shared evidence field of Eq. 3, A^(f) is the learned projection from descriptor channels to field channels, and ℓ is the comparison inside L2, defined in Eq. 7. The final loss is L = L1 + λL2 (Appendix E.9). The aggregations are printed as sums, and Table 1 marks the second-best value of each column in blue. Lines 075, 085 and 198 are now plain sentences, the citation at line 110 is complete, line 138 reads "Figure 2", and line 301 reads "summarizes the main results".

We hope these clarifications and the revised figures resolve your concerns, and we would be grateful if you would consider raising your score.

Reviewer_jKmGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies the task of passive sensing the content of a non-line-of-sight display from the diffuse illumination of the screen on the wall. The authors propose OARL, a feed-forward network with two modules built on a preliminary screen estimate. CHRRL compares local statistics and gradients between the wall observation and the preliminary estimate to predict four spatial recoverability gates (color, layout, structure, and detail), calibrated at training time against paired and nonmatching screens. OSAR progressively updates the reconstruction along a fixed color-to-detail sequence, where each stage inherits preceding states and scales its residual by the spatial gate before adding a prior correction.

The authors captured five sets of paired data (about 12,400 pairs) in a single indoor room. The authors conducted comprehensive comparisons to baselines, ablations of each component, as well as real and synthetic robustness sweeps.

Overall I think the task is an important imaging problem. The technique of combining CHRRL and OSAR is new although each component is not particularly novel. The real-world dataset of paired and structured screen content is new and a good contribution. The dataset split and mechanism experiments need more clarifications. The writing of the paper I think is overly complicating and requires improvement. I currently lean towards rejection.

Critical strengths
  • Recovering structured screen content (text, charts, UI layouts, webpages) from diffuse wall illumination is a concrete, well-motivated passive NLOS setting with clear security implications.

  • The captured real paired dataset is valuable for this NLOS task and I encourage the authors to share it with the wider imaging community.

Critical weaknesses
  • The paper does not explain how the 12,400 pairs are split into train and test sets, making it unclear if the model is inverting light transport or memorizing templates. Diffuse reflection at 0.5 m physically wipes out high frequencies, yet Figures 3 and 4 show sharp icons and bar charts with 25.56 dB PSNR on Screen, while gains outside of the "Screen" set are often small compared to the next best baseline. The authors should clarify if templates overlap between splits and provide error bars from different runs.
  • The paper frames itself as providing “initial experimental evidence” that an ordinary camera can recover hidden screen content from wall glow, but NLOS-OT (Geng et al., 2022) and NLOS-LTM (Zhang et al., 2025) already do passive screen-from-wall reconstruction with ordinary cameras, and neither is cited. Related work only covers active NLOS and generic restoration. These should be discussed and compared against.
  • The method is underspecified. The losses d, ℓ and the combiner ∆ in Eq. 13 are never defined, nor are the attention sizes, neighborhoods, κ, ρ, optimizer, schedule, or resolution.
  • The writing of the paper in general needs to be improved, as there is a lot of jargon in the text such as "Observability-Stratified Accretive Reconstruction" and continuous integral notation (∬\iint) for discrete grid operations that are hard to understand to me even as a person with an imaging background. Meanwhile, key details like the parameterization of W(0)W^{(0)} in Eq. 1, exact loss formulas in Eq. 7 and 13, and broken "?" citations (pages 3 and 21) are missing, and the 87-page appendix, while comprehensive, seems unorganized to me.
Questions for authors
  • Are the underlying wallpapers, UI windows, and chart templates strictly disjoint between the train and test splits, and how does the model perform zero-shot on an unseen category
  • Are all captures from one display/wall/room/camera? Are there quantitative results on physically different setups (Suppl. Fig. 2) rather than image-space perturbations?
  • Would the authors include a nearest-neighbor retrieval baseline over training screens? This would settle the memorization question either way.
  • Many baselines in Table 1 (DarkIR at 4 dB, SFHformer, OneRestore) are well below a plain U-Net. How were they trained and at what resolution?
2 comments

Thank you for recognizing the importance of the task and the value of the paired dataset. The split, a retrieval baseline and a second room are reported below.

  1. Source screens and templates are disjoint between training and test in all five categories (Appendix E.1, Table S6), and reconstruction stays 6 dB above retrieval of training screens. A second split of the Screen category places each template group entirely in training or entirely in test, with OARL and its strongest baselines retrained and three retrieval baselines added.
Screen, template-grouped splitPSNRSSIMLPIPS
OARL25.4720.8490.222
Restormer22.3660.7680.348
Retrieval, OARL evidence token19.4130.6870.481
Retrieval, DINOv2 features17.7350.6490.593
Retrieval, wall pixels17.6220.6410.583

OARL reproduces Table 1 within 0.086 dB and stays 6.06 dB and 0.16 SSIM above the strongest retrieval. Over three seeds on this split OARL reaches 25.463±0.073 dB against 22.350±0.096 for Restormer, and its lead on OSLock and Game also holds over three seeds (Appendix E.3). Applied directly to Browser, the Screen model reaches 16.967 dB, 3.2 dB above Restormer (Appendix E.7).

  1. NLOS-OT and NLOS-LTM are now cited and discussed in Section 2 and compared in Table 1. We trained both on all five categories under the same split and metrics. OARL leads both methods in all five categories, by 5.6 dB and 0.12 SSIM on Screen. Relative to both, OARL adds a per-location estimate of how much color, layout, structure and detail the wall supports, and this estimate gates every update.

  2. Eq. 13 and the training protocol are written out in full in Appendix E.9. d is the L1 distance averaged over valid pixels, ℓ is InfoNCE at temperature 0.10 over the paired screen and K = 8 nonmatching screens of the same category, and Δ(L1, L2) = L1 + λL2. Appendix E.9 also gives λ, ρ, the pooling block of the layout stage, the local window of the detail stage and the optimizer. Eq. 1 is followed by the definitions of its weights, offset and channel readout. Every method in Table 1 is trained per category on the same split and input, at 500×375 for Screen and 512×384 elsewhere, with the same optimizer settings and one validation-selected checkpoint.

  3. With the Table 1 weights, OARL reaches 23.107 dB in a second room, 1.97 dB above Restormer and 5.25 dB above retrieval. The second-room set has 200 Screen pairs with a different wall, camera body and camera distance, and its screens lie outside the training set. PSNR/SSIM/LPIPS are 23.107/0.797/0.316 for OARL, 21.142/0.719/0.428 for Restormer and 17.859/0.627/0.594 for retrieval (Appendix E.11).

  4. Section 3 introduces CHRRL as the gating module and OSAR as the staged refinement module, each by one plain sentence, and gives each its own figure (Figures 3 and 4). The aggregations are printed as sums over the pixel grid, the citations on pages 3 and 21 are complete, the appendix opens with a table of contents, and the controls of items 1 to 4 are in Appendix E.

We will release the dataset, splits, weights and code. We hope the grouped split, the retrieval and NLOS comparisons and the second-room results resolve your concerns, and we would be grateful if you would consider raising your score.

Dear Reviewer jKmG, thank you again for your thoughtful assessment and constructive feedback. As the discussion period is closing, we would like to kindly follow up on our response above. We have provided the template-disjoint split clarification and the additional retrieval-baseline, passive-NLOS comparison and second-room experiments, which directly target the concerns you raised.

We would also like to kindly note that Reviewer 5CYz, who raised closely related questions on the train/test separation and on transfer to a different acquisition setup, found that these same experiments address most of their concerns and make the paper more convincing, and has raised both the overall rating and the soundness score.

We hope these results have addressed your concerns as well. If so, we respectfully ask you to consider increasing your score to support our work, it is a huge motivation for us. We would also be glad to clarify any remaining points.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.