acceptodds
Withdrawn from ICLR 2027

Optical-Evidence Predictive Representation Learning for Hidden Screen Reconstruction from Diffuse Wall Reflections

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Passive hidden-screen reconstruction aims to recover the visual content of a display that lies completely outside the camera field of view from the diffuse light it casts on a visible wall. This inverse problem is challenging because indirect propagation attenuates color, luminance, and local structure at different rates, while wall reflectance, ambient illumination, and camera exposure introduce observation-dependent interference. The resulting ambiguity also encourages learned decoders to complete missing content from data priors rather than from evidence in the current observation. We propose Optical Evidence Predictive Representation Learning (OEPRL), an observation-grounded reconstruction framework that organizes wall-derived evidence before using it to correct screen-space features. Reliability-Calibrated Optical Evidence Representation (RCER) constructs an 11-channel multiscale evidence tensor, calibrates global photometric conditioning using exposure and contrast statistics, and extracts positioned local evidence tokens from informative wall regions. Observation-Consistent Optical Predictive Learning (OCPL) then predicts global and regional summaries of the selected wall evidence, establishes soft correspondences between screen and wall representations, and applies bounded global-to-local corrections to the decoder. On the Browser, Figure, Game, and Lock benchmarks, OEPRL ranks first in all 12 LPIPS, SSIM, and PSNR comparisons. On Browser, it improves over the strongest competing results by 0.0148 LPIPS, 0.0235 SSIM, and 0.714 dB PSNR. Module ablations and perturbation experiments further support the complementary contributions of evidence calibration and predictive correction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Hidden-screen reconstruction, Computational imaging, Diffuse wall reflections, Optical evidence representation, Predictive representation learning, Inverse problems

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_yYraOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether an ordinary RGB camera can recover hidden-screen content by observing the diffuse light it casts onto a visible wall. The authors propose OEPRL, which consists of two modules, RCER and OCPL: RCER organizes the mixed wall response into reliability-calibrated global evidence and positioned local tokens; OCPL uses prediction errors between wall evidence and decoder features to perform sequential global and local corrections. The authors evaluate on self-collected Browser, Figure, Game, and Lock datasets, report ranking first in all 12 LPIPS, SSIM, and PSNR comparisons, and validate the contribution of each module and a degree of robustness through ablations, occlusion, noise, illumination, viewpoint, and wall roughness experiments.

Critical strengths

The problem setting is novel, and the paper provides an initial demonstration of recovering hidden-screen content from diffuse wall light. By connecting hidden-screen recovery with an ordinary RGB camera and diffuse wall reflections, the paper expands the boundary of non-line-of-sight imaging and offers a complete system, baselines, and multi-perturbation experiments that can inform future work.

Critical weaknesses

The paper does not sufficiently justify the practical value of the problem; the scarcity of related work does not make the problem important. The authors motivate the work with hidden-screen recovery from diffuse wall reflections, but they do not provide concrete application scenarios, user needs, or threat models. The recovered information is mainly limited to brightness, dominant colors, and regional layout, while texture and fine details are largely lost; PSNR is only 17–19 dB and the visual results are blurry. If the goal is merely to show that an ordinary camera can perform some form of NLOS imaging, it is unclear whether the advantage over existing NLOS methods—such as ultrafast laser, interferometry, or polarization—is sufficient to support an independent research problem. The authors need to answer: who needs this capability, under what conditions is it useful, and what practical tasks can the recovered coarse content support? Otherwise, the problem looks more like an interesting demo than a clearly valuable research direction.

Questions for authors
  1. What are the concrete application scenarios or threat models for this capability? Who needs to recover hidden-screen content from wall reflections? Compared with existing NLOS methods such as ultrafast laser, interferometry, or polarization, is the advantage of an ordinary RGB camera sufficient to constitute an independent and important problem?

  2. Can the authors provide cross-scene, cross-wall, and cross-device generalization experiments? If all results come from the same experimental environment, how can we be sure the method has not simply memorized that environment?

  3. Does RCER’s reliability calibration have quantitative validation? How can we know it selects “reliable evidence” rather than merely “easy-to-reconstruct regions”? Can it be compared with learned uncertainty estimation or physical propagation models?

  4. Reconstruction PSNR is only 17–19 dB, mainly recovering low-frequency layout and dominant colors, with texture details largely lost. What practical use does this quality have in real scenarios? Should the authors add task-oriented metrics such as layout matching, color distribution, OCR readability, or semantic consistency?

Reviewer_wLBkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the recovery of screen content outside the camera's field of view from diffuse wall reflections captured by a standard RGB camera. The proposed OEPRL framework uses RCER to construct global and local evidence representations calibrated for exposure and contrast, and then employs OCPL to refine the decoder using global feature differences and local correspondences. With models trained separately for the Browser, Figure, Game, and Lock categories, the paper reports improvements in LPIPS, SSIM, and PSNR, and presents module ablations, wall-region intervention experiments, and experiments under various observation perturbations.

Critical strengths
  1. The problem is practically meaningful, and the acquisition approach is appealing. The paper uses an ordinary RGB camera and wall diffuse reflections to study hidden-screen recovery, demonstrating color and layout information that can be exploited from indirect illumination, which is valuable for both passive computational imaging and screen information leakage assessment.

  2. The experiments are relatively broad and attempt to verify the role of the observation area. Table 1 reports leading results on three metrics across four content categories. For example, on Browser, it improves over the strongest competitor on each metric by 0.0148 LPIPS, 0.0235 SSIM, and 0.714 dB PSNR, respectively.

Critical weaknesses
  1. The mechanism is not fully validated. The paper interprets the projection difference between two learned features as an observation-consistent prediction error, but the training objective contains no loss targeting the prediction itself. Moreover, the two rows "w/o RCER" and "w/o OCPL" in Table 2 are almost identical. For example, Browser's SSIM is exactly the same, and Game's LPIPS and PSNR are exactly the same. This suggests that without RCER, OCPL may have no input, so the contributions of the two modules are not separated.

  2. The data splits and generalization validation are insufficient to clarify the practical scope. Section 4.1 and Appendix A.2 show that the four content categories are trained separately, while the first three categories each have only 810 training pairs, and the split only explicitly excludes exact duplicate samples. It is not stated whether the split is grouped or isolated by website source, content template, or acquisition session, nor is the performance of a single fixed model on unseen categories and independent devices or rooms shown.

  3. Key results are inconsistent across tables and lack uncertainty analysis. The full model in Tables 1 and 2 reports SSIM 0.7196 on Figure and 0.6095 on Lock, whereas the corresponding values for the full model in Tables A1–A2 are 0.6989 and 0.6063. The appendix calls these values the full-model reference for each ablation group but does not explain the specific settings that cause the discrepancy. In addition, in Table 1, the best SSIM margins for Figure and Lock are only 0.0001, which cannot establish that every reported advantage is robust. Thus, the claim of "all 12 first places" is somewhat overstated.

  4. The abstract information on OpenReview is inconsistent with the actual abstract of the paper, and the paper lacks the required AI USE STATEMENT.

Questions for authors
  1. please Provide results for conventional conditional fusion with equal capacity, distillation removal, and cross-sample evidence exchange that preserves global photometric statistics, in order to distinguish the contributions of the predictive correction mechanism, teacher supervision, and input priors?

  2. Why do the full-model SSIM values differ between Tables 1–2 and Tables A1–A2? Can you provide multi-random-seed results or paired confidence intervals under a unified setting to confirm whether the small margins are reliable?

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

This paper investigates the feasibility of recovering hidden-screen content from diffuse wall illumination. It proposes OEPRL, which consists of two parts. RCER builds a global code and 16 position-tagged local tokens from low-pass/residual image statistics with hand-set exposure and contrast thresholds. OCPL applies two bounded residual corrections to the deepest decoder feature: one from a global “prediction error” between pooled projections, and one from a Sinkhorn-normalized token-to-cell correspondence. Experiments cover four content categories (Browser, Figure, Game, Lock), comparisons against about 25 restoration baselines, ablations, and an extensive appendix of robustness sweeps.

Critical strengths
  • Experiments are comprehensive.
Critical weaknesses
  • Missing prior work and overstated novelty. The first stated contribution is to “demonstrate the initial feasibility of recovering hidden-screen content from diffuse wall illumination.” This was already claimed and demonstrated by IR4Net[1]. It was demonstrated again by SGRL[2]. Neither paper is cited, discussed, or compared against. The overlap is substantial. SGRL uses identically named Browser, Figure, and Game subsets and the same 3.5 m camera-to-wall distance. IR4Net’s ReSh-WebSight and ReSh-Chart correspond to Browser and Figure. The robustness protocol in Appendix C (brightness reduction, camera–wall distance, camera position, camera orientation) mirrors IR4Net Sections 4.3–4.4.

  • The ablation results are inconsistent with the architecture. Every new module enters through tightly bounded residuals with scales of 0.08, 0.06, 0.05, and 0.03 (Eqs. 7–9, 16, 20). Yet removing either RCER or OCPL drops Browser PSNR from 19.283 to about 12.9 dB. That is 5.6 dB below the InstructIR model the network is initialized from, and below most general restoration baselines (e.g., Uformer 16.616, NAFNet 17.028). At the same time, two structurally different ablations produce nearly identical numbers. On Browser, “w/o RCER” and “w/o OCPL” have the same SSIM (0.5591) and LPIPS of 0.6897 vs. 0.6900. On Game, both have LPIPS 0.6160 and PSNR 17.30. I find these numbers suspicious, and the author presents no runnable code and data for reproduce.

  • Under additive Gaussian noise (Table 3), OEPRL on Lock scores higher at every SNR level, including 20 dB (17.04 dB PSNR), than on clean inputs (16.966 dB, Table 1). Several other appendix sweeps (brightness, wall reflectance, wall roughness) also contain perturbed conditions that exceed the clean Figure result of 19.166 dB. At the same time, many baselines are essentially invariant to the perturbations. InstructIR stays at 16.48–16.49 dB on Lock across all noise levels, and at 18.00 dB on Figure across all four roughness levels (clean: 18.002). DeepSN-Net and SFHformer are likewise constant. This pattern indicates one of two things: either the baselines produce near input-independent outputs, or the perturbation tables do not reflect actual re-evaluation. The roughness and reflectance levels (“10/20/30/40”) have no units, and the paper does not state whether conditions are physically re-captured or digitally simulated, whether test sets are shared, or whether models are retrained.

  • Baseline training quality. Several baselines appear to have failed to train. DarkIR obtains 5.047 dB PSNR and 0.0541 SSIM on Browser, and U-Net obtains 7.793 dB. Combined with the invariance noted above, this raises doubts about whether the baselines were properly tuned for the task.

[1] Physically-Guided Optical Inversion Enable Non-Contact Side-Channel Attack on Isolated Screens, ICLR 2026 [2] Bounded Hidden-Screen Reconstruction from Diffuse Wall Reflections via Survivability-Guided Representation Learning, Photonics 13(9):895, 2026

Questions for authors
  • Please elaborate the difference between your work and IR4Net and SGRL.

  • Please explain why U-Net obtains 7.793 dB? To our experience, U-Net can be a legit baseline for most of the restoration task. What makes your different?

  • Please explain why OEPRL on Lock scores higher at every SNR level under additive Gaussian noise than clean inputs? Why perturbed can exceed clean result?

  • Every new module enters through tightly bounded residuals with scales of 0.08, 0.06, 0.05, and 0.03 (Eqs. 7–9, 16, 20)? Are these module actually effective? How much do they change the feature scale?

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

Related papers

Discussion (0)

Sign in to comment.