acceptodds
Withdrawn from ICLR 2027

RECALL: Learning Content-Anchored Spectral Representation via Restoration-as-Allocation for Hidden Screen Reconstruction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Can an ordinary camera positioned behind a wall-facing display recover its con-tent from diffuse wall light alone? Propagation, scattering, occlusion, blur, andsensor integration disperse weak content cues, disrupting local screen-to-wallcorrespondence and making appearance-based capability allocation ambiguous.We present REpresentation-enabled Content-Anchored aLLocation Learning (RE-CALL), a wall-image-only framework built around Content-Anchored SpectralRepresentation (CASR) and Restoration-as-Allocation Learning (RAL). CASRorganizes patchwise spectral evidence around a sample-dependent content anchorderived from the wall image with a frozen visual encoder, combining smooth-layout and residual-detail cues. RAL learns soft spatial allocation by comparingthe pooled errors of complete reconstructions produced under local and context-augmented capability choices on the same wall observation during training. Acrossseven screen-content categories, RECALL achieves the strongest overall screen-reconstruction performance among the compared image-restoration methods, withthe highest PSNR and SSIM in every category and the lowest LPIPS in five. Acrossvaried capture conditions and screen-content scenes, it coarsely recovers hidden-screen color and brightness distributions, regional layouts, and principal contoursfrom blurred wall-light patches. These experiments demonstrate the feasibility ofobserving hidden screen content with an ordinary camera capturing only wall lightpatches.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
hidden screen reconstruction, image restoration, diffuse wall reflection, content-anchored spectral representation, restoration-as-allocation learning

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

In this work, authors propose RECALL, a method for reconstructing hidden screen content from a single image of diffuse wall light captured by an ordinary camera placed behind a wall-facing display. They frame this as an inverse problem in which optical mixing disperses screen cues many-to-many across the wall, and aim to recover coarse color, layout, and contours. Built on a two-pass encoder decoder, RECALL introduces CASR, which organizes dispersed cues around a CLIP-derived content anchor using patch-wise FFT correlation and a layout–detail gate, and RAL, which trains a router to allocate local and context-augmented corrections using soft targets derived from the errors of forced reconstructions. On about 19,700 self-collected wall–screen pairs across seven categories, RECALL outperforms 21 restoration baselines, achieving the highest PSNR and SSIM in all categories and the lowest LPIPS in five.

Critical strengths

S1. The paper tackles a new and more constrained setting than prior NLOS work, in which the camera sits behind the display and observes only the diffuse light reflected on the wall, and demonstrates on real captured data that an ordinary smartphone camera alone can coarsely recover the screen’s color and brightness distributions, layout, and principal contours.

S2. Compared with 21 restoration methods spanning unified, general, degradation-specific, and enhancement-oriented families, the proposed method achieves the highest PSNR and SSIM in all seven categories, with large margins of 6.97 dB on Software and 5.40 dB on Browser, and the qualitative results (Figure 3) show that it reproduces layouts that the other methods largely fail to recover.

Critical weaknesses

W1. The motivation for this task is insufficient. The paper does not explain why the problem is studied or where it would be applied. Given the clear security and privacy implications of recovering screen content from wall reflections, a discussion of these implications and ethical considerations is missing.

W2. Eq. 1 formulates light transport physically, but neither the network nor the training loss uses it; the model is a purely data-driven mapping learned from paired data. Comparisons with prior work on screen recovery or light-transport modeling are also missing.

W3. The ablation results are not convincing. In Table 3, the content anchor, spectral, Fourier, and detail gate ablations report identical values to four decimal places, as do the replacement experiments in Table 2 (Fourier/Context → CNN/Attention, Added Capacity). The results of w/o CASR, w/o RAL, and w/o CASR+RAL are also nearly identical, making the individual contribution of each module unclear.

W4. There is no qualitative analysis of whether the layout–detail gate in CASR and the router allocation in RAL behave as intended. All data were collected under a single geometry (0.5 m display–wall, 3.5 m camera–wall), so generalization to other wall materials, distances, angles, and displays is not validated.

W5. The notation in Eq. 2~4 is overly compressed, and most implementation details are deferred to the Appendix, making the method hard to follow and reproduce. The overall presentation needs substantial improvement.

Questions for authors

Q1. The role of Eq. 1 in the proposed method, and the reason why the physical model is not used as a prior, should be clarified.

Q2. Since wall images are heavily blurred and low in contrast, it should be demonstrated whether CLIP, trained on natural images, can extract meaningful content-related global features, with supporting analysis.

Q3. The reason why the different ablation variants in Tables 2 and 3 yield identical values to four decimal places should be explained.

Q4. Visualizations of the allocation maps are needed to show whether the router actually selects the context capability in regions with layout boundaries or dense details.

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

This paper investigates the reconstruction of hidden screen content using only diffuse wall reflections and proposes the RECALL framework. Specifically, Content-Anchored Spectral Representation (CASR) organizes spectral and multiscale features using content anchors extracted by a frozen visual encoder, while Restoration-as-Allocation Learning (RAL) learns spatial allocation based on the reconstruction errors of different restoration branches. The framework employs two restoration passes with shared weights to recover coarse colors, layouts, and principal contours of hidden screen content.

Critical strengths
  1. Meaningful research problem. Recovering hidden screen structures from weak indirect observations is a challenging task. The proposed content-conditioned representation and restoration-utility-guided allocation have reasonable motivations.
  2. Promising reported performance and extensive experiments. In Table 1, RECALL achieves the best PSNR and SSIM across all seven content categories. The supplementary material also includes spatial intervention analyses, low-light representation comparisons, various single-factor robustness tests, and routing comparisons with confidence intervals.
Critical weaknesses
  1. Numerous distinct ablation settings produce identical results. In Tables 2–3, substantially different configurations, including removing or shuffling content anchors, removing spectral processing, and replacing Fourier operations with convolution or attention, yield exactly the same results across multiple metrics and four content categories. More concerningly, in supplementary Tables A4–A5, removing utility supervision and using uniform, random, or hard-winner utility targets produce identical results up to six or seven decimal places. These highly unusual patterns raise serious concerns about the reliability and validity of the reported experimental results.
  2. Unclear fairness of baseline comparisons and attribution of the large performance gains. In Table 2, removing both CASR and RAL still achieves 20.45 dB PSNR on Software, substantially outperforming the strongest external baseline in Table 1 (14.66 dB). This suggests that a considerable portion of the reported improvement may originate from other components of the framework rather than the proposed modules. Furthermore, the paper does not sufficiently clarify whether the external baselines were retrained on the same dataset under comparable conditions, including initialization, optimizers, and training budgets.
  3. Insufficient clarity regarding dataset splits and generalization. The paper reports approximately 19,700 paired samples but only specifies the 700/100/200 training/validation/test splits for four categories. The sample distributions, training participation, and evaluation protocols for the remaining three categories remain unclear. It is also unclear whether near-duplicate samples from the same screen-content sources were excluded across splits.
Questions for authors

Could the authors explain why numerous substantially different ablation configurations produce exactly identical results? Were these variants independently implemented, trained, and evaluated? Please provide the corresponding experimental configurations and clarify how the reported results were obtained.

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

This paper presents RECAL to reconstruct hidden screen content from RGB wall reflections. It combines content-anchored spectral representation (CASR), conditioned on frozen CLIP features, with restoration-as-allocation learning (RAL), which supervises branch allocation using relative reconstruction errors. A shared-weight two-pass network reconstructs coarse colors, layouts, and contours.

Critical strengths
  1. The paper addresses an interesting and challenging inverse problem with potential relevance to optical privacy assessment.
  2. The reported results show consistent improvements across multiple content categories, with qualitative improvements in coarse layout reconstruction.
Critical weaknesses
  1. Given the severe information loss and nonlocal mixing, comparisons should include passive NLOS methods [1, 2] and strong conditional generative models conditioned on the wall observation. It is also unclear whether all baselines were retrained or fine-tuned on the same paired dataset.
  2. The ablations do not convincingly isolate the proposed components. Many distinct variants in Table 3 produce identical metrics, while different utility targets in Tables A4/A5 match to six or seven decimal places. Please explain whether these variants were independently trained and whether the modified components and losses were active. Additionally, Figure A12 reports only a 0.0037 dB clean-input gain over Hard Winner, with a confidence interval spanning zero.
  3. The dataset and generalization protocols require clarification. Explicit splits are provided for only four categories, and their relationship to the approximately 19,700 pairs remains unclear. Please specify content sources, measures to prevent similar content from crossing splits, and which capture conditions were held out during training.
  4. Figure 2 is overly dense and difficult to follow. The overview should emphasize the main processing flow, with detailed module operations presented separately. The 123-page supplement would also benefit from substantial condensation, particularly by consolidating repetitive robustness tables and highlighting the key findings.
  5. Although coarse reconstruction is the stated goal, the practical value of the recovered images remains insufficiently demonstrated. Please clarify the intended applications and evaluate whether the reconstructions reliably reveal useful screen information.
Questions for authors

Please refer to the weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.