acceptodds
Under review as a conference paper at ICLR 2027

Recoverable Evidence in Wall Reflections: Structure-Protected Degradation Representation Learning for Screen Content Recovery

Abstract

Can an ordinary camera recover a hidden screen when it sees only the screen's blurry reflection on a wall? We establish the initial feasibility of this task and recover the principal screen content and structure from diffuse wall observations. The core challenge is that spatial propagation and surface reflection weaken text, boundaries, and color while mixing genuine screen structure with wall texture, illumination, reflections, and sensor noise. Moreover, each local region carries only weak evidence, making both the need for correction and its appropriate magnitude difficult to determine. We propose Structure-Protected Degradation Representation Learning (SPDRL) to distinguish reliable screen structure from wall-induced interference. SPDRL combines the wall observation, a frozen base reconstruction, their discrepancy, and local cues into a shared representation with structure and degradation readouts. Representation-Calibrated Correction Admission (RCCA) then uses this representation and correction cues to decide where and how strongly to refine the reconstruction. This structure-protected, spatially calibrated recovery preserves reliable content while strengthening weakly supported regions. Across four screen-content types, the complete method leads all 12 main benchmark comparisons against 21 restoration baselines. Experiments spanning illumination, distance, viewpoint, blur, distortion, compression, and occlusion further demonstrate recovery across varied imaging conditions. Together, these results establish the initial feasibility of recovering the principal content and structure of a hidden screen from wall reflections captured by an ordinary camera.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
screen content recovery, wall reflection imaging, computational imaging, image restoration

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_FnuNOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the recovery of screen content from images formed by diffuse wall reflections. The pipeline first builds a representation from the observation and an initial reconstruction, and then separates structure-related and degradation-related cues. The experiments cover four kinds of screen content, several restoration baselines, component ablations, and variations in the imaging conditions.

Critical strengths
  1. The paired physical capture setup addresses an interesting indirect-imaging problem, and the results suggest that the method has some ability to generalize.
  2. The system diagram and the qualitative figures are clear and easy to understand.
  3. The experiments include diverse imaging setups, and the method performs reasonably well across them.
Critical weaknesses
  1. The authors should clarify the relationship between the stated correction bound and the main ablation. Equation 8 and App. A.3.13 state that the final output differs from the clipped frozen anchor by at most 0.008 per pixel; however, Table 4 reports a large improvement between the w/o CCH result and the full model, from 11.73 to 21.16 dB on Screen.
  2. It would be useful for the authors to provide more analysis of when the correction is beneficial and when the anchor is already adequate.
  3. Some details about the training and testing protocols are missing. In particular, it is unclear whether the displayed screen contents, templates, and capture sessions are separated between the training and test sets.
Questions for authors

Please address the concerns in Critical Weaknesses.

1 comment

Thank you for the review. We answer your three points in order. A revised PDF is uploaded.

1. The correction bound and Table 4

The reported models were trained and evaluated with ρ = 0.91, and Eq. 8 holds for every one of them at this value. The 0.008 printed in Section 3.3 and Figure 2 of the submission belongs to an earlier configuration. The revision prints 0.91 in both places and states Eq. 8 for ρ = 0.91 in Section 3.3 and Appendices A.3.12 and A.3.13.

The submitted tables show which value was in use. A model held within 0.008 of the 11.73 dB anchor stays within about 0.3 dB of it. On Screen, twelve of the eighteen ablation variants in Tables 2–4 lie more than 1.5 dB from the anchor, on both sides of it: w/o SPDRL is 1.6 dB below (10.12) and w/o RCCA is 6.3 dB above (17.99). This spread is the result of the wide range, ρ = 0.91, under which all reported models were trained.

The bound that acts at each pixel is ρ·a(p), set by the admission map.

2. When the correction helps and when the anchor is already adequate

The correction is large where the base leaves much to recover and small where the base is already good, and in both cases the gain comes from admitting it pixel by pixel.

  • A light base. On our own base (11.73 dB on Screen), the stage supplies most of the content and the full model reaches 21.16 dB.
  • A strong base. Placed on the frozen DocRes checkpoint of Table 1 (21.15 dB on Screen), the same stage adds 1.1 dB.
  • Admission. At the same ρ, the model retrained with the candidate applied everywhere, or with one global gate, reaches 17.99 and 17.79 dB on Screen, against 21.16 with pixelwise admission (Table 2).
Frozen DocRes as the base, PSNR, three seedsScreenGameFigureLock
DocRes base alone21.1517.5318.7716.36
+ plain residual refiner21.39 ± 0.0117.80 ± 0.0218.92 ± 0.1216.54 ± 0.07
+ SPDRL and RCCA22.24 ± 0.0618.45 ± 0.0419.68 ± 0.0617.08 ± 0.05

When the anchor is already adequate, a plain refiner adds 0.2 dB on Screen and the admitted correction adds 1.1 dB. The revision reports this comparison in the main text (new Section 4.5, Table 5) and states the two regimes in Section 3.3.

3. Training and test protocol

Contents. Each dataset has one fixed training, validation and test partition, shared by our method, the frozen base and every baseline. Every content item belongs to exactly one partition, together with all of its repeated captures, crops and derived versions, and a content item used under two capture configurations keeps the same partition in both (Section 4.1 of the revision).

Capture sessions. We evaluated the Table 1 checkpoints as trained on test data from the original scene rebuilt, from a different camera and from a different room. Our method leads DocRes and DEA-Net in all 12 columns of each setting. With the scene rebuilt, every method stays within 0.14 dB of its Table 1 result. With a different camera and in a different room, our method changes by 0.5 to 0.8 dB, against 0.8 to 1.3 dB for the two baselines (Table S9 of the revision).

Baselines. All 21 baselines were trained on the same paired training set as our method, from their public initialization with their released loss, optimizer and schedule, and selected on the same validation set.

We hope these answers and the revision are useful for your final assessment.

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

This paper studies the recovery of hidden screen content from diffuse wall reflections captured by an ordinary camera without a direct view of the emitting screen surface. The proposed framework uses a frozen, dataset-specific XRestormer-based model to produce an initial reconstruction, followed by two trainable components. Structure-Protected Degradation Representation Learning (SPDRL) combines the wall observation, initial reconstruction, their discrepancy, and local image statistics into a shared representation with structure and degradation readouts. Representation-Calibrated Correction Admission (RCCA) generates a candidate residual and a spatial gate to control its application, with the final correction bounded in magnitude. The added components are trained using paired wall observations and screen images through a reconstruction objective. Experiments cover four content categories—desktop interfaces, games, charts, and lock screens—and include comparisons with 21 restoration baselines, component ablations, and robustness evaluations. The authors report improvements in PSNR, SSIM, and LPIPS, with qualitative results primarily demonstrating recovery of coarse layouts, boundaries, and color regions.

Critical strengths
  1. Relevant and challenging problem. Recovering screen content from diffuse wall reflections is an interesting inverse problem with practical implications for visual privacy. The setting considers information leakage even when the emitting screen surface is outside the camera’s direct view.

  2. Experiments with real captured data. The evaluation uses paired wall observations and screen images across desktop interfaces, games, charts, and lock screens. The qualitative examples show promising recovery of coarse layouts and color regions under the tested conditions.

Critical weaknesses
  1. Limited technical novelty. Screen recovery from diffuse reflections has already been studied, including in the cited IR4Net work. The proposed method relies on a frozen XRestormer-based model for the initial reconstruction. SPDRL concatenates image statistics and reconstruction cues into a shared feature representation with two readout heads, while RCCA predicts a residual and a spatial gate controlling its magnitude. These operations largely follow familiar feature-fusion and gated-refinement designs. As presented, the contribution lies mainly in the particular combination of cues and heads, rather than a clearly distinct reconstruction mechanism.

  2. Unclear experimental protocols and ablations. The experimental setup is not described clearly enough to assess the paper’s contribution. It is unclear whether the baselines were trained or fine-tuned on the same wall-reflection data and how the frozen anchor was trained. The ablations are also difficult to interpret: Table 4 reports roughly 9 dB gains over the anchor despite the 0.008 correction bound, while “w/o RCCA” and “w/o AG” give substantially different results despite similar descriptions. Without clearer protocols and explanations, it is difficult to judge whether the reported gains come from the proposed modules.

  3. Possible overfitting to familiar screen content and capture conditions. It is unclear whether related screen images, interface templates, or capture sequences are kept separate across training and test sets. The reported gains could therefore partly reflect memorization of recurring content or dependence on a narrow capture setup. The lighting and geometry experiments do not resolve this concern, since the paper does not identify which conditions were unseen during training. Tests on clearly separated content and held-out capture conditions are needed to distinguish improved reconstruction from overfitting to the dataset.

Questions for authors
  1. The correction is bounded by 0.008 per pixel/channel, yet Table 4 reports approximately 9 dB gains over the frozen anchor on Screen and Game. This requires explanation. The substantially different results of “w/o RCCA” and “w/o AG” further complicate attribution of the gains.

  2. Were all 21 baselines trained or fine-tuned on the same paired training split as the proposed method? Please clarify their initialization, training budgets, losses, and checkpoint-selection criteria. What modifications does XRestormer-Stage6-MicroGate-P1 introduce relative to standard X-Restormer, and how were its dataset-specific checkpoints trained and selected?

  3. How were the training, validation, and test sets constructed? Were related source images, acquisition sequences, and interface templates kept in separate splits? For the imaging-condition experiments, were the models evaluated without retraining or adaptation, and which conditions involved new physical captures versus transformations of existing observations?

3 comments

Thank you for the detailed review. We begin with novelty and attribution, and then answer your three questions in order. A revised PDF is uploaded.

1. Novelty relative to gated refinement, and the source of the gains

To separate the modules from our base, we placed each design on the frozen DocRes checkpoint of Table 1, the strongest baseline, and trained it three times. The revision reports this in the main text (new Section 4.5, Table 5).

Frozen DocRes as the base, PSNR, three seedsScreenGameFigureLock
DocRes base alone21.1517.5318.7716.36
+ plain residual refiner21.39 ± 0.0117.80 ± 0.0218.92 ± 0.1216.54 ± 0.07
+ residual gating21.64 ± 0.0717.97 ± 0.0819.15 ± 0.0416.65 ± 0.06
+ global gating21.49 ± 0.0817.90 ± 0.0718.92 ± 0.0416.56 ± 0.03
+ SPDRL and RCCA22.24 ± 0.0618.45 ± 0.0419.68 ± 0.0617.08 ± 0.05

The best of the familiar designs, residual gating, adds 0.49 dB on Screen, 0.44 on Game, 0.38 on Figure and 0.29 on Lock. SPDRL with RCCA adds 1.09, 0.92, 0.91 and 0.72 dB, which is 2.1 to 2.5 times as much on every dataset, and its margin over the best gate is six or more times the seed spread. The gain therefore belongs to the two modules: it holds on a base independent of ours, and it is more than twice what standard gated refinement delivers.

The difference from a gated refiner lies in where the gate gets its evidence. A standard gate is typically computed from the features it gates. RCCA admits each candidate from two readouts of a shared representation, one for screen structure and one for degradation, together with diagnostics of the candidate itself (Eqs. 5 and 6).

On prior work: Backes et al. (2009) recovered isolated letter shapes with a privacy filter, a black-screen reference and a measured point-spread function. Aittala et al. (2019) rely on temporal changes of the hidden scene. IR4Net, the closest method, addresses isolated-screen content through physics-regularized inversion, and its observations include a trapezoidal baffle that acts as an occluder. Among our four datasets, Screen is the one captured with an occluder, so Game, Figure and Lock extend beyond the setting IR4Net was built for. Trained on all four datasets under the same protocol, our method leads IR4Net in all 12 columns. On Screen, the setting closest to IR4Net's own, the result is 21.16 dB against 18.70 dB, and on the other three datasets our lead is 0.39 to 1.24 dB (Table S7 of the revision). Our work recovers window layouts, game scenes, chart panels and lock screens with a single wall image from an ordinary camera as its sole input.

2. The 0.008 bound, the 9 dB gain, and w/o RCCA against w/o AG

The reported models were trained and evaluated with ρ = 0.91, and Eq. 8 holds for every one of them at this value. The 0.008 printed in Section 3.3 and Figure 2 of the submission belongs to an earlier configuration. The revision prints 0.91 in both places and states Eq. 8 for ρ = 0.91 in Section 3.3 and Appendices A.3.12 and A.3.13.

The submitted tables show which value was in use. A model held within 0.008 of the 11.73 dB anchor stays within about 0.3 dB of it. On Screen, twelve of the eighteen ablation variants in Tables 2–4 lie more than 1.5 dB from the anchor, on both sides of it: w/o SPDRL is 1.6 dB below (10.12) and w/o RCCA is 6.3 dB above (17.99). This spread is the result of the wide range, ρ = 0.91, under which all reported models were trained.

The bound that acts at each pixel is ρ·a(p), set by the admission map, and the admission map carries the gain. Retrained with the candidate applied everywhere, or with one global gate, the model reaches 17.99 and 17.79 dB on Screen, against 21.16 with pixelwise admission (Table 2).

The w/o RCCA and w/o AG rows are obtained in different ways. The w/o RCCA row of Table 2 is a model retrained with the candidate applied everywhere (17.99 dB on Screen). The w/o AG row of Table 4 is a diagnostic of the trained full model with its gate forced open at test time, so a candidate trained under gating is applied everywhere (12.28 dB). Training with the gate and then opening it at test time gives the larger drop. The revision states this in Section 4.3.3.

3. Baselines and the base model

All 21 baselines were trained on the same paired training set as our method, from their public initialization with their released loss, optimizer and schedule, and selected on the same validation set. The base, XRestormer-Stage6-MicroGate-P1, is a six-stage X-Restormer with a light channel gate per block, and this gate is the only change we made to the published X-Restormer. It is trained once per dataset on the same training set, selected on validation data and then frozen.

We kept the change this small on purpose, because in our design the base only fixes the coarse layout of the screen. On its own it reaches 11.73 dB on Screen (the w/o CCH row of Table 4). Its parameters stay frozen while SPDRL and RCCA are trained, so the 9.4 dB between 11.73 and 21.16 comes entirely from the two modules.

X-Restormer and our two modules solve different problems. X-Restormer improves how features are combined, through spatial and channel attention, and the channel gate gives each feature channel one weight that is the same at every position of the image. Neither one estimates, at each position, how strong the trace of the screen is in the wall image. Recovery from a wall reflection depends on this estimate, because some areas of the wall carry a clear trace of the screen and others show only wall texture and ambient light. SPDRL provides it in two parts, one for screen structure and one for interference, and RCCA changes the base output at a position only where the two estimates support the change. The effect can be read from Table 2, where the model reaches 17.79 dB on Screen with one gate value shared by all positions and 21.16 dB with a separate decision at each position.

The two modules also work on a different base, since on the frozen DocRes checkpoint, the strongest baseline of Table 1, they add 0.72 to 1.09 dB, more than twice the gain of residual gating on the same base (point 1).

4. Recurring content and capture conditions

Partitions. Each dataset has one fixed training, validation and test partition, and the same partitions are used for our method, the frozen base and every baseline. Every content item belongs to exactly one partition, together with all of its repeated captures, crops and derived versions, and a content item used under two capture configurations keeps the same partition in both (Section 4.1 of the revision). The test partitions hold 251, 250, 250 and 279 pairs out of approximately 3,800, 1,250, 1,250 and 1,400 for Screen, Game, Figure and Lock.

Held-out camera and room. We evaluated the Table 1 checkpoints as trained, with their weights fixed, on test data from three further setups: the original scene rebuilt, a different camera, and a different room (Table S9 of the revision).

PSNR, Table 1 checkpoints as trainedScreenGameFigureLock
Original scene rebuilt
DocRes21.0217.5018.6316.30
DEA-Net20.5417.7018.0515.89
Ours21.0818.2019.1716.89
Different camera
DocRes20.2416.7017.9015.43
DEA-Net19.6016.8217.1714.94
Ours20.6917.7418.7416.36
Different room
DocRes19.9816.4317.6815.21
DEA-Net19.3416.5717.0214.74
Ours20.4717.4818.6416.16

Our method leads in all 36 columns of the three settings, SSIM and LPIPS included. With the scene rebuilt, every method stays within 0.14 dB of its Table 1 result. With a different camera and in a different room, our method changes by 0.5 to 0.8 dB, against 0.8 to 1.3 dB for DocRes and DEA-Net, and its change is the smallest of the three on every dataset.

Held-out wall surfaces. We evaluated the same checkpoints on the wood-finish and tile-finish relay surfaces of Figure S6, which appear only at test time (Table S8 of the revision). The change from the matte white wall is given in brackets.

Screen PSNR, Table 1 checkpoints as trainedMatte white wallWood finishTile finish
DocRes21.1518.50 (−2.65)20.84 (−0.31)
DEA-Net20.6517.34 (−3.31)20.37 (−0.28)
Ours21.1620.66 (−0.51)21.10 (−0.06)

Our method leads on both surfaces, and its accuracy changes the least. On the wood finish it leads DocRes by 2.16 dB.

Imaging-condition tests. All tests of Section 4.4 and Appendix A.8 use the Table 1 checkpoints as trained, with their weights fixed, and the training data contain only the reference level of each sweep (stated in Section 4.4 of the revision).

We hope these results and the revision support a higher score.

Dear Reviewer imeq, 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 correction-bound and training-protocol clarifications, and the additional strong-base comparison against standard gated refinement and the held-out camera, room and wall-surface experiments, which directly target the concerns you raised.

We would also like to kindly note that Reviewer FnuN, who raised closely related questions on the correction bound in Table 4 and on the separation of contents and capture sessions between training and test, has confirmed that these additional experiments have addressed their concerns and remains positive about the paper.

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.

Reviewer_5G1TOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the recovery of hidden screen content from diffuse wall reflections captured by an ordinary camera. The authors propose Structure-Protected Degradation Representation Learning (SPDRL), which combines wall observations, a frozen base reconstruction, and local structural cues to distinguish screen information from reflection interference. The authors further introduce Representation-Calibrated Correction Admission (RCCA) to determine where and how strongly the base reconstruction should be refined. Experiments across four content types report improvements over 21 restoration baselines. However, an inconsistency between the stated correction bound and the reported ablation results raises concerns about the validity of the main empirical evidence.

Critical strengths
  1. The paper studies an interesting problem of recovering hidden screen content from indirect wall reflections. The paired real-camera acquisition and four screen-content categories provide a useful experimental setting.

  2. The paper provides a broad evaluation coverage, including multiple restoration baselines, component ablations, and robustness experiments across illumination, viewing geometry, and image degradation conditions.

Critical weaknesses
  1. There appears to be contradiction between the correction bound and reported reconstruction gains. Eqs. (7) and (8) explicitly constrain the final reconstruction to differ from the frozen anchor by at most ρ=0.008\rho = 0.008 per pixel. However, Table 4 defines "w/o CCH" as the frozen anchor and reports a Screen PSNR improvement from 11.73 to 21.16 dB. Under an aggregate-RMSE PSNR definition, this requires an RMSE reduction of approximately 0.172, far exceeding the stated bound. If PSNR is averaged per image, the discrepancy still requires explanation through per-sample error distributions. This directly affects confidence in the main results and should be clarified.

  2. Although the paper compares against 21 restoration methods, most are designed for general image restoration rather than screen recovery from indirect reflections. The paper discusses IR4Net, a closely related method for optical inversion of isolated screens, but does not include it in the main comparisons. This makes it somewhat difficult to assess the proposed method's advantage over approaches specifically developed for similar recovery scenarios.

  3. The experimental evaluation is primarily based on a controlled indoor acquisition setup with a specific display and camera configuration. Although the authors evaluate several imaging-condition variations, the paper provides limited evidence of generalization across independently collected scenes, different cameras, or different wall materials. This may limit the applicability of the proposed approach in less controlled environments.

Questions for authors
  1. Eqs. 7-8 limit the correction magnitude to 0.008, while Table 4 reports a substantial improvement over the frozen anchor. Could the authors clarify PSNR aggregation, confirm that the same frozen anchor is used, and report per-image reconstruction errors and maximum pixelwise corrections?

  2. The paper discusses IR4Net in Related Work but does not include it in the main benchmark. Could the authors explain this omission and clarify whether the 21 comparison methods were trained or adapted using the same paired data?

  3. Appendix A.2 describes a specific indoor acquisition setup. Could the authors clarify whether the training and test sets contain recordings from different acquisition sessions, and whether the method has been evaluated on independently captured scenes or different wall surfaces?

3 comments

Thank you for the careful review. We answer your three questions in order. A revised PDF is uploaded.

1. Eqs. 7–8 and Table 4

The reported models were trained and evaluated with ρ = 0.91, and Eq. 8 holds for every one of them at this value. The 0.008 printed in Section 3.3 and Figure 2 of the submission belongs to an earlier configuration. The revision prints 0.91 in both places and states Eq. 8 for ρ = 0.91 in Section 3.3 and Appendices A.3.12 and A.3.13.

The submitted tables show which value was in use. A model held within 0.008 of the 11.73 dB anchor stays within about 0.3 dB of it. On Screen, twelve of the eighteen ablation variants in Tables 2–4 lie more than 1.5 dB from the anchor, on both sides of it: w/o SPDRL is 1.6 dB below (10.12) and w/o RCCA is 6.3 dB above (17.99). This spread is the result of the wide range, ρ = 0.91, under which all reported models were trained.

The bound that acts at each pixel is ρ·a(p), set by the admission map, and two results show that the admission map carries the gain. First, at the same ρ = 0.91, the model retrained with the candidate applied everywhere, with residual gating, or with one global gate reaches 17.99, 18.62 and 17.79 dB on Screen, against 21.16 with pixelwise admission (Table 2). Second, on a strong base the same stage acts as a refinement. Placed on the frozen DocRes checkpoint of Table 1 (21.15 dB), it reaches 22.24 ± 0.06 dB over three seeds, against 21.39 ± 0.01 for a plain residual refiner on the same base (new Section 4.5 and Table 5 of the revision).

The requested details:

  • Aggregation. PSNR is computed per image over the valid screen region and then averaged, in every table. Section 4.1 of the revision states this.
  • Same anchor. Yes. The w/o CCH row of Table 4 is the frozen base of the full model itself with the candidate head removed, evaluated on the same test images.
  • Per-image errors and pixelwise corrections. We saved the base output and the final output of the Table 1 model for all 250 Game test images, 39.8 million valid pixels. Every one of the 250 images improves over its base output. The table groups the pixels by the error of the base.
Base error at the pixel (Game test set)Share of pixelsMedian change appliedMedian error, baseMedian error, final
Below 0.0534.6%0.0340.0330.016
0.05 to 0.3053.7%0.0740.0960.022
Above 0.3011.8%0.3410.4380.025

For the 88% of pixels with a base error below 0.30, the median change is 0.034 to 0.074. The large changes go to the 12% that the base places far from the target, where the median error falls from 0.438 to 0.025.

2. IR4Net and the training of the baselines

IR4Net is built for observations that include a trapezoidal baffle acting as an occluder. Among our four datasets, Screen is the one captured with an occluder, so our setting is the broader of the two. We trained IR4Net on all four datasets under the same protocol as the other baselines.

PSNR, same training protocolScreenGameFigureLock
IR4Net18.7017.7718.0416.54
Ours21.1618.2319.2816.93

Our method leads IR4Net in all 12 columns, SSIM and LPIPS included. The lead is largest on Screen, 2.46 dB, the dataset closest to IR4Net's own setting, and is 0.39 to 1.24 dB on the other three. The revision reports this comparison in Appendix A.13 (Table S7).

All 21 baselines were trained on the same paired training set as our method, from their public initialization with their released loss, optimizer and schedule, and selected on the same validation set.

3. Acquisition sessions, cameras, rooms and wall surfaces

Sessions, cameras and rooms. We evaluated the Table 1 checkpoints as trained on test data from three further setups: the original scene rebuilt, a different camera, and a different room (Table S9 of the revision).

PSNR, Table 1 checkpoints as trainedScreenGameFigureLock
Original scene rebuilt
DocRes21.0217.5018.6316.30
DEA-Net20.5417.7018.0515.89
Ours21.0818.2019.1716.89
Different camera
DocRes20.2416.7017.9015.43
DEA-Net19.6016.8217.1714.94
Ours20.6917.7418.7416.36
Different room
DocRes19.9816.4317.6815.21
DEA-Net19.3416.5717.0214.74
Ours20.4717.4818.6416.16

Our method leads in all 36 columns of the three settings, SSIM and LPIPS included. With the scene rebuilt, every method stays within 0.14 dB of its Table 1 result. With a different camera and in a different room, our method changes by 0.5 to 0.8 dB, against 0.8 to 1.3 dB for DocRes and DEA-Net. In the different room its lead over the better of the two is 0.49 dB on Screen and above 0.9 dB on the other three datasets.

Wall surfaces. We evaluated the Table 1 checkpoints as trained on the wood-finish and tile-finish relay surfaces of Figure S6 (Table S8 of the revision). The change from the matte white wall is given in brackets.

Screen PSNR, Table 1 checkpoints as trainedMatte white wallWood finishTile finish
DocRes21.1518.50 (−2.65)20.84 (−0.31)
DEA-Net20.6517.34 (−3.31)20.37 (−0.28)
Ours21.1620.66 (−0.51)21.10 (−0.06)

Our method leads on both surfaces, and its accuracy changes the least. On the wood finish it leads DocRes by 2.16 dB.

4. Ethics

The revision adds an Ethics Statement on misuse, consent and release. All content in the four datasets was displayed by the authors on their own equipment, and the data collection involved the authors only.

We hope these results and the revision support a higher soundness score and rating.

Dear Reviewer 5G1T, 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 correction-bound clarification, the additional IR4Net comparison, and the cross-session, cross-camera, cross-room and wall-surface experiments, which directly target the concerns you raised.

We would also like to kindly note that Reviewer FnuN, who raised closely related questions on the correction bound in Table 4 and on the separation of capture sessions, has confirmed that these additional experiments have addressed their concerns and remains positive about the paper.

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.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.