acceptodds
Withdrawn from ICLR 2027

AGTRL: Axis-Guided Trusted Representation Learning for Passive Hidden-Screen Recovery

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

When a screen is completely hidden from direct view, its emitted light can still diffuse onto a nearby wall and encode information about the displayed content. We investigate whether an ordinary camera can recover that content solely by observing these wall light traces. The central challenge is that propagation and diffuse reflection compress screen appearance into low-contrast brightness and color variations, mix contributions from different screen regions, and superimpose screen-related traces with wall texture, shadows, exposure variation, and sensor noise. We propose Axis-Guided Trusted Representation Learning (AGTRL) to recover hidden-screen content by learning both wall-to-screen correspondence and trace reliability. Axis-Guided Representation Learning (AGRL) organizes dispersed wall features along channel and spatial axes to form a screen-oriented representation. Embedded TraceTrust arbitration combines local continuity, artifact suspicion, and a learned gate to prioritize credible screen traces and suppress pseudo traces during reconstruction. We conduct extensive experiments across four screen-content families and under changes in distance, brightness, viewpoint, noise, occlusion, and wall appearance. AGTRL outperforms 20 mainstream image reconstruction methods in all 12 dataset–metric combinations and robustly recovers the principal color distributions, regional positions, page layouts, and structural contours of hidden screens. These results demonstrate the initial feasibility of observing and recovering hidden-screen content using an ordinary camera aimed only at wall light traces.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
passive hidden-screen recovery, indirect imaging, image reconstruction, representation learning, wall-light observations

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_VTGdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes AGTRL for recovering coarse hidden-screen content from a single RGB observation of its diffuse light on a wall. The method uses an encoder–decoder with two bottleneck components: AGRL forms a normalized convolutional feature update, and TraceTrust weights this update using local continuity, local deviations, and a learned gate. The network is trained with paired screen targets using an L1 reconstruction loss. Evaluation covers four screen-content datasets, comparisons with image-restoration methods, component ablations, and robustness experiments.

Critical strengths

The paper reports consistent improvements across four datasets, with a substantial gain on Screen2000. Compared with the strongest baseline in Table 1, PSNR improves from 23.44 to 24.71 dB, SSIM from 0.7990 to 0.8371, and LPIPS from 0.3261 to 0.2587. These results provide useful evidence that the proposed design can improve reconstruction in the evaluated setting.

The component evaluation is reasonably extensive. Table 2 removes or replaces the main components, while the supplementary experiments examine finer design choices and compare the combined module with standard attention mechanisms and a parameter-matched convolutional alternative in a shared U-Net. These controls provide support for the architectural contribution beyond the headline comparison.

Critical weaknesses
  1. The baseline selection does not adequately establish the claimed advance over existing passive reconstruction methods. Section 2 explicitly discusses closely related approaches, including NLOS-OT, NLOS-LTM, and IR⁴Net, yet none is included in Table 1, and their exclusion is not explained. Comparing against a large collection of general restoration architectures leaves the most relevant task-specific comparison unresolved. The authors should evaluate an appropriate existing passive reconstruction method on the same paired data, or explain the concrete acquisition, supervision, or implementation constraints preventing such a comparison. Without this, the claim of state-of-the-art performance rests on an incomplete comparison set.

  2. The baseline adaptation and training protocol is insufficiently documented to assess the fairness of Table 1. Section 4.1 specifies paired splits and states that a separate model is trained for each dataset, but does not explain how the 20 baseline methods were initialized, adapted, and trained for wall-to-screen reconstruction. The reported MAE/AdamW settings appear in the AGTRL description and are not explicitly stated to apply to the baselines. Since many compared methods were originally developed for different restoration tasks, their task-specific training and model selection are essential to interpreting the results. Without the training schedules, tuning procedure, and checkpoint-selection criteria, it is difficult to determine whether the reported advantages reflect the proposed design or differences in how the methods were adapted and optimized.

Questions for authors
  1. Why are the closely related passive reconstruction methods discussed in Section 2 absent from Table 1? Please provide a comparison with an appropriate method on the same paired data, or explain the specific constraints preventing it.

  2. Were all 20 baselines trained or fine-tuned on the stated wall–screen training splits? Please clarify their initialization, task-specific modifications, training schedules, tuning procedure, and checkpoint selection.

Reviewer_s6B9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes AGTRL for passive hidden-screen recovery using wall-reflected light captured by an ordinary camera. The method introduces two components: AGRL, which organizes channel and spatial features for screen reconstruction, and TraceTrust, which weights feature updates based on local continuity, artifact suspicion, and a learned gate. Experiments on four datasets demonstrate its reconstruction performance under various observation conditions.

Critical strengths
  1. The paper studies an interesting and practically relevant problem of recovering hidden-screen content from indirect optical observations.
  2. The experiments are extensive, including comparisons with 20 reconstruction methods, component ablations, and robustness evaluations under various conditions.
Critical weaknesses

Insufficiently Substantiated Methodological Contribution. The paper's central methodological contribution is not sufficiently established. Although AGRL and TraceTrust are presented as task-specific solutions for hidden-screen recovery, their designs rely largely on conventional operations, with limited theoretical or physical justification for their connection to the underlying optical problem. Consequently, the paper does not convincingly demonstrate a substantial methodological advance beyond combining existing techniques, which is my primary concern regarding its acceptance.

Questions for authors
  1. Limited methodological novelty and insufficient physics-based justification. AGRL mainly consists of channel normalization, pointwise convolutions, and depthwise convolutions, while TraceTrust combines local statistics with a learnable gating mechanism. These operations are relatively conventional network designs, and the substantive novelty introduced by their combination remains unclear. Furthermore, the paper only provides an abstract description of the imaging process through x=ψ(y;η). The subsequent network design does not explicitly incorporate light propagation principles, nor does it sufficiently explain how these specific operations address the spatial mixing of reflected light. Many mathematical formulations primarily describe standard network computations, without providing deeper theoretical analysis of the proposed method.

  2. Insufficient justification and validation of TraceTrust. TraceTrust implicitly assumes that locally continuous features are more reliable, whereas features with larger local deviations are more likely to represent interference or artifacts. However, genuine reflected screen signals may also contain meaningful local variations, which could therefore be incorrectly suppressed. The existing experiments mainly demonstrate improvements in reconstruction quality, without directly validating whether the estimated trust weights can effectively distinguish genuine screen signals from interference. The authors are encouraged to provide visualizations of the trust maps and controlled experiments analyzing the relationship between trust weights and genuine signal or interference regions.

  3. Unexplained variations in ablation results. Table 2 shows that removing AGRL or TraceTrust reduces PSNR by over 5 dB on Screen2000. However, Supplementary Table 24 reports only a 0.135 dB improvement over the Identity baseline when adding both modules to a shared U-Net. Although the configurations differ, this large discrepancy warrants clarification. The authors should explain the architectural and training differences and assess the dependence of module effectiveness on the backbone.

  4. Insufficient explanation of experimental trends and metric inconsistencies. Supplementary Table 26 shows that reducing wall-image contrast unexpectedly improves PSNR and SSIM, despite potentially weakening reflected signals. The authors attribute this to robustness against weak signals, but contrast reduction may also suppress wall textures and interference. This trend therefore requires further explanation. Additionally, Supplementary Tables 24 and 25 report identical PSNR (11.986 dB) on ScreenFigure but noticeably different SSIM and LPIPS values. The authors should clarify whether the U-Net architectures, checkpoints, and evaluation settings are consistent, and explain these differences.

  5. Overly dense figure and table presentation. Figure 2 combines the acquisition setting, network architecture, module details, mathematical expressions, and experimental illustrations, making the main design difficult to follow. Some tables, particularly Table 3, are also excessively dense. Simplifying the architecture figure and improving table readability would enhance the presentation.

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

This paper investigates passive hidden-screen recovery from screen light reflected onto a nearby wall and recorded by an ordinary RGB camera. It proposes Axis-Guided Trusted Representation Learning (AGTRL), an encoder–decoder framework combining channel normalization and spatial aggregation through AGRL with a TraceTrust mechanism that weights latent residual updates using local continuity, artifact-suspicion statistics, and a learned gate. The paper evaluates the method on four screen-content datasets under two acquisition configurations, reporting the best PSNR, SSIM, and LPIPS among 20 reconstruction methods across all four datasets. Extensive supplementary experiments investigate component alternatives and sensitivity to noise, masking, geometric transformations, and wall-appearance variations. The reported reconstructions primarily recover coarse color distributions, region arrangements, and screen layouts.

Critical strengths
  1. A worthwhile inverse-imaging problem with encouraging empirical results. Recovering hidden-screen information from diffuse wall reflections using an ordinary camera is practically relevant to computational imaging and visual privacy. Table 1 reports the best results across all 12 dataset–metric combinations. The Screen2000 improvement is particularly notable: PSNR increases from 23.44 to 24.71 dB over UniProcessor, SSIM increases from 0.7990 to 0.8371, and LPIPS decreases from 0.3261 to 0.2587. These results provide promising evidence that the proposed approach can improve reconstruction under the evaluated acquisition conditions.

  2. Substantial investigation of the proposed components. The experimental evidence extends beyond simple component removal. Table 2 evaluates alternative residual branches and gates, while the supplementary material examines individual projection, aggregation, and weighting choices. Supplementary Table 24 also compares representation modules within a shared U-Net, including SE, CBAM, non-local processing, FFC, and a parameter-matched convolutional alternative. These comparisons are valuable because they provide evidence that the proposed combination contributes to reconstruction quality, rather than relying exclusively on comparisons between unrelated full architectures.

Critical weaknesses
  1. The evaluation does not adequately establish how much recovered structure is supported by the optical measurements rather than learned content priors. Section 4.1 specifies non-overlapping paired splits and trains a separate model for each dataset, but does not sufficiently establish separation by source content, interface template, game sequence, or acquisition session. Figure 6 demonstrates sensitivity to wall-input masking, but this does not distinguish faithful recovery from predictions combining weak measurement cues with recurring content patterns. Content-disjoint evaluation and controls such as spatially averaged wall-color inputs or nearest-neighbor retrieval would help resolve this ambiguity. Furthermore, several supplementary robustness results concern digital transformations or wall-appearance proxies; these support perturbation tolerance but do not by themselves establish transfer to new physical light-transport conditions. This is consequential because the central contribution depends on identifying what screen information is recoverable from the measurements and under which acquisition conditions. This concern is not an allegation of data leakage.

  2. The evidence supports the usefulness of the architecture more clearly than its claimed interpretation as correspondence learning and trace-reliability estimation. AGRL combines channel normalization, pointwise projections, depthwise spatial convolution, and GELU, while TraceTrust relies on smoothness-related latent statistics. Such statistics may be effective, but they do not inherently distinguish genuine screen information from nuisance features. In Section 3.3, whenever the local deviation reaches the spatial mean plus epsilon, the artifact-suspicion term saturates at one, making the proposal weight zero regardless of the learned gate. The identity path remains available, but the paper does not establish whether this suppression preferentially removes artifacts or also suppresses informative structure. The ablations show reconstruction benefits, yet direct validation of the reliability interpretation is missing. In addition, the relationship between the main AGTRL results and the shared-U-Net experiments is insufficiently explained—for example, Screen2000 PSNR is 24.706 dB in Table 2 versus 9.097 dB for U-Net with AGRL + TraceTrust in Supplementary Table 24. These different settings are not necessarily inconsistent, but clearer training and architectural specifications are needed to attribute the gains. Together, these issues limit confidence in the claimed methodological insight and currently lead me toward weak rejection.

Questions for authors
  1. Dataset independence and measurement information. Are the training and test sets disjoint by source image, interface template, game sequence, and acquisition session, or only by paired sample? What evidence shows that the recovered spatial structure exceeds what can be predicted from recurring content patterns or global wall-color statistics? Content-disjoint results or simple information-restricted controls would materially strengthen my assessment.

  2. TraceTrust interpretation. What fraction of latent locations have a saturated artifact-suspicion value of one, and how does this fraction change under corruption? Is there evidence that the resulting trust weights distinguish corrupted regions from informative screen-related structure, rather than primarily favoring smooth features? Please clarify whether “trust” denotes a validated reliability estimate or a heuristic reconstruction weight.

  3. Experimental comparability and attribution. What backbone, training, preprocessing, and checkpoint-selection differences explain the Screen2000 results of 24.706 dB in Table 2 and 9.097 dB for U-Net with AGRL + TraceTrust in Supplementary Table 24? How were the principal baselines adapted and trained, and is the large Screen2000 ablation improvement stable across independent training runs?

  4. Scope of physical generalization. Which quantitative robustness tests use newly captured observations, and which transform existing RGB images? The appendix explicitly identifies actual acquisition distances but separately describes image-transformation and wall-appearance-proxy tests. Please specify which physical conditions were held out from training and what conclusions about transfer to new acquisition configurations are supported by those measurements.

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

The topic of this work is interesting, as it studies passive hidden-screen recovery in which an RGB camera observes only the weak light pattern a hidden screen casts on a nearby diffuse wall, and the displayed content is reconstructed from this wall observation alone. It proposes AGTRL, a U-Net-based network trained end-to-end on paired wall-screen images, with two bottleneck modules: AGRL, which employs channel normalization and depthwise spatial aggregation forming a residual proposal, and TraceTrust, which uses a trust map built from hand-crafted continuity/suspicion statistics and a learned gate that weights the residual. Experiments on four self-collected datasets compare against 20 restoration networks.

Critical strengths
  • This research topic is to some degree interesting;

  • Paired data collection is a promising contribution;

  • Experiments are extensive, and result analysis is comprehensive.

Critical weaknesses
  • My first concern is the baseline selection. This work compares with 20 baselines of image restoration, which mainly model diverse degradations rather than optical processes, which seems unconvincing to me. The authors are suggested to select baselines from condition-based or physically aware generators rather than only selecting from the restoration field.

  • It is unclear whether the model learns optical inversion or memorizes content patterns. The method contains no physics-based modeling of wall-to-screen light transport, each model is trained and tested on the same scene and rig, and there is no evidence that test screen contents differ substantially from training contents. Thus, the authors are suggested to quantify train-test target similarity (nearest-neighbor SSIM/LPIPS of each test target against the training set

  • The contribution of the architecture in this work is insufficient. AGRL seems to be channel normalization plus pointwise-depthwise-pointwise aggregation, and TraceTrust is a CBAM-style spatial gate built from hand-crafted smoothness statistics. And the motivation for the combination is not convincing to me.

Questions for authors
  • Does a trained model transfer to a new wall/geometry, or is a per-scene paired collection required?

  • Compare with at least one indirect-imaging baseline (Aittala 2019 or Zheng 2026) or combine it with your current baselines.

  • Please also state how many runs were attempted per configuration and confirm that identical training protocols and checkpoint-selection rules were used for the full model and all ablation variants. Some results in Table 2 seem unconvincing that no matter you remove what component the performance drop is 5 dB.

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

Related papers

Discussion (0)

Sign in to comment.