acceptodds
Withdrawn from ICLR 2027

CoRER: Coherent Relay Evidence Restoration for Initial Hidden-Human Contour Observation in Corner-Occluded Scenes

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Corner occlusion completely hides a person from direct view, while a visible relay wall carries weak reflected light from the hidden person. We study whether an RGB relay-wall spot image captured by an ordinary camera supports an initial observation of the hidden person's basic region and main contour. The task is challenging because local wall appearance dominates target-associated variation after diffuse relay transport, and contour support is distributed across decoder scales. We present the Coherent Relay Evidence Restoration Model (CoRER), whose Relay Diffuse-Coherent Representation (RDCR) organizes weak front-end evidence with projected local features, scale-separated smoothing, disagreement cues, and a bounded residual update. Its Scale-Coherent Relay-State Restoration (SCRR) measures smooth and detail agreement at decoder skip fusions, organizes multiscale support and reliability, and recalibrates preliminary foreground-background logits to restore the main contour. An explicit positional calibration supplies a bounded shape prior before the final argmax readout. Extensive comparisons, ablation studies, robustness evaluations, and qualitative results validate CoRER's ability to recover the hidden person's basic region and main contour from faint reflected-light spots on the relay wall. Across four real corner scenes in the ReflectHuman benchmark involving different people, corner structures, illumination conditions, and scene settings, CoRER outperforms representative hidden-scene perception and image segmentation methods overall and recovers the basic region and main contour from faint reflected-light spots on the relay wall captured by an ordinary camera, demonstrating the feasibility of this initial contour observation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
image segmentation, contour segmentation, binary segmentation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GXtWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CoRER reconstructs the silhouette of a person hidden around a corner using a single RGB image containing weak indirect reflections from a wall. Its architecture combines RDCR and SCRR to achieve this reconstruction. The study also incorporates the ReflectHuman dataset, a real-world dataset acquired in four corner environments, and evaluates the method through baseline comparison, component ablation, and robustness experiments. This task presents a challenging scenario for non-line-of-sight (NLOS) imaging.

Critical strengths
  1. ReflectHuman supports research on a challenging real-world perception task. Capturing faint wall reflections in multi-corner environments provides useful experimental data for recovering hidden human contours. This dataset allows researchers to examine non-line-of-sight (NLOS) reconstruction under real-world imaging conditions and compare the performance of different methods on common tasks.

  2. The architecture is tailored to a specific problem and validated from multiple experimental perspectives. The separation of RDCR and SCRR provides a clear modular structure for the reconstruction process. In addition to overall performance comparisons, the authors investigated individual components through ablation experiments, robustness checks, boundary-sensitive metrics, and diagnostic tests. These analyses help characterize the network's behavior.

Critical weaknesses
  1. The distinctions between CoRER and closely related methods are not adequately clarified. This paper omits discussion of two directly related studies: SLD-Net (CVPR 2026), entitled “Hidden Person Detection Based on Similarity-Consistent Likelihood Diffusion from Wall Reflections,” and DiProxB-Net (IEEE Transactions on Multimedia, 2026), entitled “Recovering Occluded Human Boundaries from Single-Frame Indirect Observations.” Both studies explore methods for hidden person perception or boundary reconstruction using indirect wall observations. Their implementations differ from CoRER, but they overlap significantly in perceptual assumptions, task definitions, and research motivations. The comparison with DiProxB-Net is particularly important because it also reports results on the ReflectHuman dataset. This paper needs to more clearly articulate CoRER's contributions beyond these methods, supplemented with targeted numerical comparisons and visual examples under similar conditions.

  2. Some interpretations and evaluation conclusions lack sufficient justification. RDCR appears to assess consistency between features across different scales, but the use of "coherence" fails to clearly distinguish this computational property from optical coherence. Standard RGB images record intensity rather than directly measuring optical phase, thus any physical interpretation requires careful articulation. Similarly, phase jitter analysis fails to adequately link its image spatial transformation to the underlying measurement process. The evaluation scheme also suffers from uncertainty regarding test independence. The paper should explain the distribution of scenes, acquisition sequences, and participants during the training, validation, and testing phases, and clearly indicate which reported results used reserved data. Furthermore, current evidence is insufficient to distinguish between reconstructions based on information-rich wall reflections and predictions primarily shaped by learned prior information about human contours.

Questions for authors
  1. What are the specific technical advantages of CoRER compared to SLD-Net and DiProxB-Net? Given the implementation and data, could the authors compare these methods using matched inputs, evaluation protocols, and representative visual results?

  2. How are the training, validation, and test sets constructed? Please explain whether the acquisition sequences, environments, or subjects are shared between the sets, and describe measures taken to prevent accidental overlap.

  3. Does "coherence" refer to feature consistency, optical properties, or other quantities? Please define the phase jitter transform and explain the hypotheses it tests. Using control variables such as empty scenes, known changes in human position, or selectively occluded wall observations can also help determine the dependence of the prediction results on the measured reflected signal.

  4. Could the authors provide the complete training objective and loss weights, define the calibration operator (B_j), and describe each step used in the inference process? For baseline methods developed for transient measurements or LiDAR data, please explain how they are converted to single-frame RGB input and compare how their different sensing assumptions are considered.

Reviewer_NQn4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies whether an ordinary RGB camera can recover the region and main contour of a person completely hidden around a corner, using only faint reflections on a flashlight-illuminated relay wall. The proposed CoRER model maps a single relay-wall image to a binary human mask. It has two main components: RDCR, which compares short- and long-range smoothed features at the front end and uses their disagreement to weight a residual update, and SCRR, which pools smooth-support, detail, agreement and reliability cues across decoder stages to recalibrate the output logits, followed by an explicit positional shape prior. The authors evaluate on ReflectHuman, four real indoor corner scenes with 24,781 relay-wall images paired with masks annotated from a synchronized direct-view camera. CoRER achieves the best Dice, IoU and Boundary-F among 17 baselines in all four scenes, and the paper reports ablations, robustness tests (brightness, camera angle, distance, wall material, synthetic noise) and representation diagnostics.

Critical strengths
  1. Ambitious problem with a real paired dataset. Recovering a hidden person's silhouette from a single ordinary RGB image of a flashlight-illuminated relay wall, without transient sensors, would be practically meaningful. The paper collects about 24.8K real relay-wall images across four indoor scenes with masks annotated from a synchronized direct-view camera, which could become a useful benchmark if released.

  2. Broad experimental coverage. The evaluation includes 17 baselines from three families, component removals and replacements, several real-capture robustness sweeps, and diagnostics of intermediate representations, giving a wide view of the model's behavior.

Critical weaknesses
  1. Generalization of the central feasibility claim is not established. The paper's main claim is that faint wall reflections carry usable information about the hidden person's shape. However, the data are acquired with a camera recording at 30 fps, and the paper does not explain how the fixed training–validation split was made. If frames were split at random, near-duplicate frames would appear in both sets. Checkpoints are also selected by validation Dice and results are reported on the same validation set, with a single seed. This matters because the results are surprising: in Fig. 3, CoRER recovers fine pose details (e.g., a crouching posture or gaps between the arms) from what appears to be a nearly uniform bright spot, while across five relay-wall materials Dice varies by only 0.0024 and HD95 by 0.16 px (Table A12), and on Reflect-Room Dice changes by less than 0.001 across the brightness sweep. These observations would be consistent with memorization of near-duplicate frames as well as with genuine signal recovery. Without a session-disjoint evaluation and controls that isolate the contribution of the relay-wall signal, the reported results do not conclusively establish whether the recovered contours arise from genuine indirect-light evidence rather than scene-specific priors or temporal correlations.The qualitative examples are also curated, so they cannot settle this question on their own.

  2. Baselines and method description are not adequately specified. All 17 baselines fall within a narrow Dice range (0.63–0.76) with low Boundary-F (0.10–0.24), and several were designed for entirely different measurements: PlatoNeRF uses two-bounce lidar, NLOST and TRT-NLOS use transient data, and RISE targets RIS-based wireless imaging. The paper does not describe how they were adapted to a single RGB image, beyond training all models with the same 50-epoch recipe. In addition, the method is described mostly in abstract terms, with operators such as Ψr\Psi_r, QrQ_r, FiF_i and BjB_j not concretely specified even in App. B, and there is no code or reproducibility statement, so the results are difficult to reproduce or verify.

Questions for authors
  1. How was the training–validation split constructed? Are training and validation frames drawn from disjoint recording sessions (or disjoint time segments), or were frames split at random from the same videos? If the latter, please report results on a session-disjoint split.

  2. What Dice, IoU and Boundary-F are obtained by (a) predicting the mean training mask of each scene, (b) CoRER with blanked or randomly shuffled inputs, and (c) the positional calibration prior alone? These controls would directly show how much of the performance comes from the relay-wall signal.

  3. How were the real-capture robustness experiments performed, given that the reference contour is held fixed across brightness levels and wall materials? Please explain why performance is nearly invariant to wall material, and describe how the transient-, lidar- and RF-based baselines were adapted and trained on single RGB images.

Reviewer_8YnBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether an ordinary RGB camera can recover the basic region and main contour of a person hidden around a corner by observing faint reflections on a visible relay wall. It proposes CoRER, consisting of RDCR, which organizes relay evidence using responses at different spatial supports, and SCRR, which restores contour information across decoder scales. The experiments use four real indoor scenes from the ReflectHuman benchmark, and compare CoRER with hidden-scene and segmentation baselines. The paper reports that CoRER achieves the highest Dice, IoU, and Boundary-F scores among the compared methods in each scene, with additional ablation and robustness analyses.

Critical strengths
  1. The paper studies an interesting and potentially useful sensing problem that inferring a coarse hidden-person contour from relay-wall observations captured with an ordinary RGB camera.

  2. The evaluation covers four real scenes and includes baseline comparisons, component ablations, and robustness analyses, providing more evidence than a single benchmark score.

Critical weaknesses
  1. The evaluation protocol, as described in the submission, uses fixed training and validation partitions and selects checkpoints by validation Dice, but I could not identify a separate held-out test set. It is also unclear whether images from the same person, acquisition sequence, or closely related conditions may appear in both partitions. This limits the strength of the generalization claims.

  2. The experiments do not yet clearly rule out that the model exploits scene-specific, illumination-specific, or acquisition-specific correlations rather than target-related reflected-light evidence. More controlled tests, such as holding out people, acquisition sequences, or scenes, would strengthen the central claim.

  3. The paper does not appear to provide sufficient information about dataset access, code availability, or the procedures used.

Questions for authors
  1. How were the training and validation partitions constructed? Are people, recording sessions, and acquisition sequences separated across the partitions?

  2. Do the authors have results on an independently held-out test set, or using leave-one-scene/person/session-out evaluation?

  3. What controls establish that predictions depend on target-related relay evidence rather than scene identity, wall appearance, or illumination patterns? For example, have the authors tested target-absent inputs or controlled changes in target position while keeping the scene and acquisition conditions fixed?

  4. Will the ReflectHuman data and code be made available?

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

Related papers

Discussion (0)

Sign in to comment.