acceptodds
Withdrawn from ICLR 2027

SplitIR: Operator-Splitting Reconstruction of Hidden Screen Content from Indirect Radiance

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Can an ordinary camera recover the content of a hidden screen from the light it casts on a visible wall? Diffuse scattering weakens and mixes the screen's structural cues, obscuring the correspondence between wall evidence and the layout to be recovered. Successive refinements can then accumulate incompatible corrections in an unrestricted correction space and suppress the weak high-frequency cues that carry fine screen structure. We propose Operator-Splitting Reconstruction from Indirect Radiance (SplitIR), which separates reconstruction into potential assimilation and resolvent correction. The Potential Assimilation Module (PAM) uses a learned positive metric and bounded potential-guided transport to organize wall evidence into a coherent layout state. The Resolvent Correction Module (RCM) then analytically adjusts candidate corrections through a metric-aware proximal step and projects them into stage-conditioned feasible correction sets for local refinement. Across Screen, Chart, Password, and Website, SplitIR achieves the best PSNR, SSIM, and LPIPS among 22 representative baselines and reconstructs screen layouts, chart structures, and character patterns from diffuse wall observations. Ablations support the two-operator design, while controlled luminance and Gaussian-noise experiments show a persistent comparative advantage as observation quality degrades. Together, these results establish the initial feasibility of recovering hidden screen content with an ordinary camera observing wall radiance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
indirect radiance, hidden content reconstruction, computational imaging, inverse imaging, operator splitting, image restoration, representation alignment, proximal optimization

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

This paper incorporates convex‑optimization principles into neural‑network modules to address the challenge from the ill‑posed non‑line‑of‑sight (NLOS) reconstruction problem. By introducing the Potential Assimilation Module (PAM) and the Resolvent Correction Module (RCM), the proposed SplitIR method outperforms existing CNN-based, Transformer-based and specialized NLOS imaging algorithms across multiple quantitative metrics on four distinct datasets. Extensive ablation studies and robustness evaluations further validate the effectiveness and limitations of the designed modules.

Critical strengths
  1. The Operator‑Splitting idea is conceptually novel. By combining convex optimization with deep‑learning architecture, the model interpretability is improved.
  2. The paper adopts a comprehensive set of comparison baselines. The ablation studies and robustness evaluations are thorough and well‑executed.
Critical weaknesses
  1. The learned metric, center, coordinate-wise radius in this paper are dynamically predicted by neural networks during the forward propagation. While such a paradigm is common for plug‑and‑play hybrid deep‑unfolding architectures, it raises interpretability concerns. The paper provides no validation to verify whether these network‑predicted parameters are physically meaningful or accurate. Moreover, lacking online correction mechanisms, these learned predictions may produce invalid outputs when test samples deviate from the training distribution, e.g., under mismatched scene geometry or noise levels.
  2. Five competing baselines (DRS, Flatland, NLOS‑OT, TransiT, PromptIR) yield identical quantitative metrics on the Password dataset. The provided qualitative output of NLOS‑OT is an all‑white reconstruction, which appears abnormal.
Questions for authors
  1. How do these adaptive parameters (metric, center, radius) behave in practice? I suggest the authors to add supplementary statistics or visualizations to verify or interprete the parameter design. Moreover, as the training loss consists purely of reconstruction‑oriented objectives, there are no additional loss terms to regularize the values of these dynamically predicted parameters. It is also suggested to investigate the benefits of introducing appropriate prior‑based penalty terms.
  2. I strongly suggest the authors to clarify the baseline metrics. Does the designed experimental setup appear to be so specialized that these baselines cannot adapt to, or does the training process simply collapse when reproducing these comparison methods?
Reviewer_taEROpenReview ↗
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 SplitIR, a feed-forward network with two modules. PAM is a bounded, gated residual update driven by a learned diagonal-weighted difference between layout-state features and observation features. RCM takes decoder residual proposals and blends each with a predicted “center” using a closed-form diagonal proximal step, then clips the result to a predicted per-coordinate box.

The authors captured four sets of paired data in a single room. The authors conducted comprehensive comparison to baselines, ablations of each component as well as synthetic luminance and noise sweeps.

Overall I think the task is an important imaging problem. The technique of combining PAM and RCM is new although each components is not particularly novel. The real-world dateset of paired and structured screen content is new and a good contribution. The experiments and ablation needs 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, page layouts) from diffuse wall illumination is a concrete, well-motivated passive NLOS setting with clear security implications.
  • The capture 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 presents itself as the first demonstration that an ordinary camera can recover hidden screen content from wall radiance. NLOS-OT (Geng et al., 2022) and NLOS-LTM (Zhang et al., 2025) already do passive screen-from-wall reconstruction. Although the paper cites both methods, what’s actually new is the content categories (charts, text, webpages) and the dataset, which is a narrower contribution.
  • The gains of the performance on the real experiments outside of the "screen set" seems small if compared to the next best baseline, with even 0.25 dB psnr, and this is with checkpoint selecting. The authors should provide more details on the performance, including error bars from different runs.
  • The writing of the paper in general needs to be improved, there are a lot of jargon in the text such as "Operator splitting", "Learned positive metric", that are hard to understand to me even as a person with imaging background. Also the appendix, while comprehensive, seems unorganized to me.
  • The paper needs to discuss and compare against deep unrolling methods (like ISTA-Net, ADMM-net) which to me are similar to PAM and RCM.
Questions for authors
  • Why do all ten ablated variants on Screen cluster at 22 dB versus 27.16 dB psnr for the full model, with removing a single stage embedding having similar performance as removing the entire module?
  • Many baselines have similarly bad PSNR (8.76 dB), can the authors clarify they are fail similarly spectacularly?
  • What does this work establish beyond NLOS-OT and NLOS-LTM, which already reconstruct hidden displays from ordinary-camera wall observations?
  • A comparison against an unrolled baseline that uses the transport model in Eq. 1 would show whether the formulation contributes beyond the architecture itself.
Reviewer_2gRgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies hidden-screen reconstruction from RGB observations of the light reflected on a visible wall. The proposed SplitIR separates reconstruction into a Potential Assimilation Module, which integrates wall evidence into a shared layout representation, and a Resolvent Correction Module, which refines image predictions through a learned proximal adjustment and box projection.

Critical strengths
  1. The problem is interesting and practically relevant. Recovering screen content from indirect wall radiance using an ordinary camera is a challenging inverse problem, and the paper provides acquisition photographs and a detailed formulation of the imaging setting.
  2. SplitIR achieves the best reported PSNR, SSIM, and LPIPS across all four datasets in Table 1. The improvement on Screen is substantial, reaching 27.16 dB compared with 24.10 dB for NLOS-OT. The qualitative examples also show improvements in layout and character structure.
  3. The experimental analysis is extensive. In addition to PAM and RCM ablations, Appendix C.4.3 includes real acquisition changes in illumination, wall material, camera motion, viewing geometry, and lens configuration. These experiments provide useful evidence beyond synthetic noise and brightness perturbations.
Critical weaknesses

1.Several baselines have identical Password scores in Table 1, including DRS, Flatland, NLOS-OT, TransiT, and PromptIR. The authors should explain why these methods converge to identical predictions. 2.The method overview Fig.2 is messy and hard to understanding. 3.The benefit of the operator formulation over a learned gated refinement network needs clearer evidence. Equations (10)-(11) implement coordinate-wise interpolation followed by clipping with learned parameters. Appendix B establishes useful properties with the predicted geometry fixed, but these properties do not establish descent or stability of the complete network when the geometry changes with the reconstruction. Table 2 includes residual and gated-residual replacements, which is helpful, but their parameter counts and computational costs are not reported. Capacity-matched comparisons would make the source of the improvement clearer. 4.The paper uses separate dataset-specific checkpoints for Screen, Chart, Password, and Website. These results therefore do not establish that a single model can reconstruct arbitrary screen content or handle an unseen content category. 5.The physical experiments reveal substantial sensitivity to camera geometry as shown in Appendix C.4.3.

Questions for authors
  1. Can the main comparisons be repeated with separate training, validation, and test partitions? How are related screen images, webpage templates, and acquisition sessions separated across these partitions?
  2. What are the character error rate and exact recovery rate on Password? How much readable text or correct chart information is recovered on Website and Chart?

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

Related papers

Discussion (0)

Sign in to comment.