acceptodds
Under review as a conference paper at ICLR 2027

DarkOcc: Illumination-Robust Indoor Occupancy Prediction

Abstract

Existing RGB-based occupancy models typically assume well-illuminated inputs and suffer substantial performance degradation under low-light conditions. Image enhancement only partially addresses this issue, while jointly training on normal- and low-light images can compromise performance under normal illumination. We present DarkOcc, a framework for robust indoor occupancy prediction across illumination conditions. DarkOcc first learns illumination-robust visual features through multi-level feature adaptation to improve occupancy prediction under low-light conditions. A geometry-guided occupancy completion module then combines these features with base depth and voxel predictions to recover missing occupied structure, addressing occupancy incompleteness exacerbated by low-light degradation. On Occ-ScanNet-mini, DarkOcc improves Geo-IoU over GPOcc trained on mixed-light data (MLA) by 5.57 pp under severe low light and 5.13 pp under normal illumination. Further experiments demonstrate its effectiveness in embodied occupancy prediction. We also construct a real-world indoor occupancy benchmark covering normal and low-light conditions. Under low illumination, DarkOcc improves Geo-IoU over MLA by 9.76 pp.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Indoor 3D Occupancy Prediction, Low-Light Robustness, Feature Adaptation, Geometry-Guided Occupancy Repair, Synthetic-to-Real Generalization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_agiBOpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
2(+1)
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies RGB-based indoor occupancy prediction under low light. It builds on GPOcc and adds two modules. Multi-Level Feature Adaptation (MFA) puts residual adapters on four VGGT encoder levels, Geometry-Guided Occupancy Completion (GOC) is a small 3D network that adds residual scores to voxels the base model predicts as free. Training uses normal-light images plus synthetic low-light versions (Dark1–Dark3). The paper also collects a small real-world indoor low-light test set. DarkOcc reports gains over mixed-light-trained GPOcc (MLA) on Occ-ScanNet, EmbodiedOcc-ScanNet, and the real-data set.

Critical strengths

The direction is practically useful. Robots often operate in poorly lit indoor spaces, and current occupancy models clearly break down there. Studying this setting and building a real low-light test set is a meaningful step.

Critical weaknesses
  1. The main comparisons should use the same experimental setting. Only MLA, IAB, and DarkOcc share a training setup. All other baselines (EmbodiedOcc, SplatSSC, FLM-Occ, original GPOcc, etc.) use original checkpoints trained on normal light only.

  2. MFA feels too hand-crafted (MFA's illumination conditioning comes from manually designed gates. There's no ablation showing these gates matter compared with a plain adapter. And GOC isn't really related to illumination, it works more like a general completion module.

  3. Insufficient discussion and comparison with low-light perception work. LIAR (nighttime occupancy prediction) and DepthDark (low-light depth estimation) are cited but neither discussed in depth nor compared experimentally. There is also a body of low-light recognition work in detection and depth estimation, these should be discussed, and compared by plugging them into GPOcc.

Questions for authors

Could you compare all methods under the same training setting?

Could you show MFA's illumination gates actually matter versus a plain adapter, and explain how GOC is related to illumination?

Could you discuss and compare with LIAR, DepthDark, and other low-light perception methods?

5 comments
AuthorsOpenReview ↗

Response to Reviewer #agiB (1/3)

Thank you for recognizing the relevance of our work to indoor robotics and the value of our real-world low-light test set. We have carefully considered your comments and would like to address them as below.
We invite you to revisit the revised manuscript, where we have clarified the presentation throughout, redrawn the DarkOcc framework in Figure 2, and added a flowchart in Figure 3 showing how our team constructed and verified the real-world occupancy labels.


W1 & Q1 [Comparing methods under matched training illumination]

Table 1 of the revised manuscript reports results for both the original and mixed-light checkpoints on Occ-ScanNet-mini. The low-light gap between DarkOcc and the normal-light-only checkpoints also reflects differences in training illumination. We retrained EmbodiedOcc, SplatSSC, and FLM-Occ under the same mixed-light training conditions. The GPOcc row in Table R1 corresponds to mixed-light augmentation (MLA). It retains the original GPOcc architecture and uses the same training illumination.

Table R1. Occupancy prediction after mixed-light training for the four baselines included in this comparison. Entries show Geo-IoU / mIoU (%).

Mixed-light trainingOriginalDark1Dark2Dark3Real normal lightReal low light
EmbodiedOcc55.75 / 47.2254.49 / 45.9151.83 / 42.6845.86 / 36.6024.17 / 14.2119.57 / 6.24
SplatSSC63.06 / 51.2661.40 / 49.5058.29 / 46.1851.62 / 39.6627.82 / 15.0219.72 / 5.08
FLM-Occ57.53 / 49.5157.55 / 49.5557.46 / 49.4056.97 / 48.6320.42 / 12.2515.37 / 4.31
GPOcc59.07 / 54.5558.76 / 54.0158.46 / 53.5556.39 / 50.8237.08 / 22.0719.36 / 11.51

On the real low-light test set, mixed-light training raised EmbodiedOcc’s Geo-IoU from 14.40% with its original checkpoint to 19.57%. Among the three baselines we retrained, SplatSSC had the highest Geo-IoU (19.72%) and EmbodiedOcc the highest mIoU (6.24%). On the same test set, DarkOcc achieved 29.12% Geo-IoU and 15.56% mIoU, exceeding all four mixed-light baselines on both metrics. Tables 1 and 3 now identify the mixed-light checkpoints for Occ-ScanNet-mini and the real indoor dataset, respectively.


W2 & Q2 [What illumination gating adds to a standard adapter and how GOC addresses missed occupancy in low light]

Thank you for raising these questions. To isolate the effect of gating, we use the same residual adapter architecture, update cap, and VGGT fine-tuning protocol in both variants. We also compare geometry-guided occupancy completion (GOC) with a 3D network of similar capacity.


Please see the subsequent Authors' Rebuttal (2/3) for remaining responses.

AuthorsOpenReview ↗

Response to Reviewer #agiB (2/3)

Continued: Authors’ Rebuttal (2/3)


(1) What does illumination gating add to a standard adapter? The gate sets the update magnitude using local brightness, the fraction of dark pixels, contrast, and mean image brightness. The residual adapter learns the correction itself. Both variants use residual adapters at the same four encoder stages, apply the same bound on the update magnitude, and omit GOC. Only the illumination gate changes (Table R2).

Table R2. Illumination-gating ablation with the same residual adapters and update bound. Metrics are reported as Geo-IoU / mIoU. The gain Δ is the score with full MFA minus the score without the gate, in pp (percentage points).

Test conditionNo light gate (%)Full MFA (%)Gain from gating (pp)
Occ-ScanNet-mini · Original61.12 / 56.4861.47 / 56.61+0.35 / +0.13
Occ-ScanNet-mini · Dark357.59 / 52.2159.02 / 52.91+1.43 / +0.70
Real indoor · normal light41.61 / 23.4642.86 / 23.93+1.25 / +0.47
Real indoor · low light21.22 / 12.7623.97 / 13.87+2.75 / +1.11

On paired frames, adding the gate improved Geo-IoU by 0.35 and 1.43 pp under normal light and Dark3, respectively. The corresponding gains on the real normal- and low-light groups were 1.25 and 2.75 pp, with mIoU also improving in both groups. As the same frames were progressively darkened from Original to Dark3, the mean update magnitude at the fifth encoder layer rose across the four lighting conditions, reaching 0.54%, 1.48%, 2.68%, and 2.99% of the input feature norm, respectively. The gated adapter thus makes small corrections under normal light and progressively larger ones as illumination falls. Compared with the ungated adapter, it improves Geo-IoU under both conditions, with larger gains in low light.

How does GOC address missed occupancy in low light? GOC was motivated by the paired experiment in Figure 1(c). Even after multi-level feature adaptation (MFA), as the same Occ-ScanNet-mini test frame grew darker, the model missed more of the occupied voxels it correctly identified under normal light. GOC therefore uses predicted depth and the surrounding 3D structure to recover occupied voxels missed by the base model. The ablation revealed an unexpected benefit. GOC also improved predictions under normal light, where the base model still misses some occupied voxels. This additional benefit serves our goal of improving low-light robustness while preserving normal-light accuracy. Comparing full MFA without GOC (Table R2) with full GOC (Table R3), adding GOC raised Geo-IoU by 3.52 pp on the real normal-light dataset and by 5.15 pp on the real low-light dataset. The capacity-matched comparison below tests how much of that gain a standard 3D refinement network can achieve.

Table R3 compares full GOC with a standard 3D residual network of comparable capacity and a variant that retains only the local branch.

Table R3. Full GOC, standard 3D refinement, and the local-branch-only variant on the real indoor data. Metrics are reported as Geo-IoU / mIoU. The gain Δ is the full GOC score minus the score of the variant in that row, in pp.

ConfigurationNormal light (%)Normal light gain (pp)Real low light (%)Real low light gain (pp)
Full GOC46.38 / 25.53—29.12 / 15.56—
Capacity-matched generic 3D residual network in place of GOC46.14 / 25.42+0.24 / +0.1128.16 / 15.08+0.96 / +0.48
Remove coarse branch, retain local branch45.88 / 25.35+0.50 / +0.1827.73 / 14.98+1.39 / +0.58

Standard 3D refinement. Under low light, the standard 3D residual network raised the base model’s Geo-IoU from 23.97% to 28.16%. At comparable capacity, full GOC reached 29.12%, outperforming standard refinement by 0.96 pp in Geo-IoU and 0.48 pp in mIoU. The corresponding margins under normal light were 0.24 and 0.11 pp.

Coarse-scale 3D context. The local branch operates at the original voxel resolution to preserve boundary detail. When a predicted structure is fragmented, the coarse-scale branch can draw on the surrounding 3D layout beyond a voxel’s immediate neighborhood. Adding the coarse-scale branch improved Geo-IoU under both lighting conditions, with a larger gain under low light (+1.39 pp) than under normal light (+0.50 pp).


Please see the subsequent Authors' Rebuttal (3/3) for remaining responses.

AuthorsOpenReview ↗

Response to Reviewer #agiB (3/3)

Continued: Authors’ Rebuttal (3/3)


W3 & Q3 [Comparison with LIAR, DepthDark, and other low-light methods and their applicability to GPOcc]

Thank you for suggesting a broader comparison with related work. To our knowledge, DarkOcc is the first method specifically developed for RGB-based indoor 3D semantic occupancy prediction under low light. Our goal is to improve robustness to low light while maintaining accuracy under normal light. LIAR predicts outdoor occupancy from multiple cameras. The other methods address monocular depth estimation or 2D semantic segmentation. Table R4 compares the task settings, input views, and adaptation mechanisms of four methods for low-light perception.

Table R4. Task settings and feature adaptation in low-light perception methods. A check mark (✓) indicates that the original paper reports the setting or mechanism. A dash (—) means it was not reported.

MethodYear-VenueOutdoor occupancy predictionIndoor occupancy predictionSingle-view RGB inputIllumination gating at multiple encoder stagesGeometry-Guided Occupancy Completion (GOC)
LIAR2025-NeurIPS✓————
DepthDark2025-ACM MM——✓——
NightFormer2024-ECCV——✓——
DarkOcc—✓✓✓✓

What each method optimizes. LIAR uses illumination to guide 2D feature sampling and refine bird’s-eye-view (BEV) features in nighttime driving scenes. DepthDark uses grayscale guidance and multi-scale features to improve per-pixel depth estimation. NightFormer enhances texture and matches reliable pixels to semantic prototypes for nighttime segmentation. DarkOcc uses occupancy supervision to adapt encoder features across lighting conditions and combines predicted depth with 3D context to recover occupied voxels missed after feature adaptation.

Transferring LIAR components to GPOcc. LIAR constructs vehicle-centered BEV features from six synchronized surround-view images. GPOcc takes one RGB view per time step and represents the scene with 3D Gaussians. These differences in input format and scene representation make it difficult to train LIAR directly on Occ-ScanNet in its original configuration. We evaluated two GPOcc configurations with LIAR components. In the first, we integrated LIAR’s SLLIE module directly into GPOcc. In the second, we trained GPOcc with both SLLIE and 2D-IGS. Table R5 reports both configurations under all four lighting conditions on Occ-ScanNet-mini. In the revised manuscript, Table 19 reports the LIAR transfer results, and Table 1 provides the GPOcc and DarkOcc reference scores.

In separate experiments, we plugged depth predictions from UniDepth V1 and DetAny3D into GPOcc and observed no gain in either Geo-IoU or semantic mIoU under any of the four lighting conditions. The lack of publicly available official implementations makes it difficult to transfer DepthDark and NightFormer to GPOcc.

Table R5. Transfer of LIAR components to GPOcc on Occ-ScanNet-mini. Results are Geo-IoU / semantic mIoU (%). For GPOcc variants, Year-Venue identifies the source of the added component.

MethodYear-VenueOriginalDark1Dark2Dark3
GPOcc2026-CVPR63.16 / 55.8560.71 / 53.6352.84 / 45.6135.65 / 28.58
GPOcc + LIAR (SLLIE)2025-NeurIPS62.98 / 55.6260.23 / 53.0651.98 / 44.8135.29 / 28.40
GPOcc + LIAR (SLLIE + 2D-IGS, trained)2025-NeurIPS63.48 / 56.0761.58 / 54.2655.74 / 47.5942.49 / 33.19

Results on Occ-ScanNet-mini. SLLIE alone scores slightly below GPOcc in all four conditions. The trained SLLIE + 2D-IGS configuration improves both metrics at every light level, with the largest gains on Dark3 (+6.84 pp in Geo-IoU and +4.61 pp in semantic mIoU). DarkOcc achieves higher Geo-IoU and semantic mIoU than both LIAR configurations under all four lighting conditions.


We hope these responses address your questions and welcome further discussion.

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

This paper investigates illumination-robust indoor 3D semantic occupancy prediction, highlighting degraded visual representations and increased occupancy incompleteness under low-light conditions. The paper proposes DarkOcc, which combines Multi-Level Feature Adaptation (MFA) to improve visual representations and Geometry-Guided Occupancy Completion (GOC) to recover missing occupied structures. Experiments on synthetic and real low-light datasets demonstrate improved occupancy prediction performance under both normal- and low-light conditions.

Critical strengths

The paper addresses the practical challenge of illumination-robust indoor occupancy prediction. The analysis in Fig. 1 provides evidence that low-light degradation exacerbates occupancy incompleteness even after multi-level feature adaptation.

Critical weaknesses
  1. Insufficient analysis of the relationship and performance trade-off between normal- and low-light conditions. A central motivation of this paper is to improve robustness to low illumination without sacrificing accuracy under normal illumination. Although the proposed MFA incorporates illumination-conditioned modulation and bounded adaptation, the paper does not sufficiently explain why this trade-off occurs or demonstrate how these mechanisms specifically address it. The experiments mainly report performance improvements across illumination conditions, without explicitly analyzing the relationship between normal- and low-light predictions. Consequently, it remains unclear whether the proposed framework effectively resolves the stated trade-off or primarily achieves general improvements in occupancy prediction accuracy.

  2. Incomplete justification of the underlying cause of occupancy incompleteness. In the Introduction section, it shows that stronger low-light degradation increases the number of occupied voxels misclassified as free space, even after feature adaptation. However, this observation alone does not establish insufficient volumetric reasoning as the primary cause of these errors. Other factors, such as inaccurate depth estimation, degraded Gaussian occupancy predictions, or unreliable occupancy classification, may also contribute to occupancy omissions. Although the proposed Geometry-Guided Occupancy Completion (GOC) improves prediction accuracy, the current experiments do not sufficiently demonstrate that the improvements specifically arise from enhanced volumetric reasoning rather than general voxel refinement. The paper lacks targeted ablation studies to establish the connection between the identified failure mode and the proposed solution.

  3. Limited Comparisons with Low-Light Occupancy Prediction Methods The comparative experiments in Table 1 mainly include general indoor occupancy prediction methods, none of which are specifically designed for low-light conditions. Although Table 3 provides comparisons with low-light image enhancement methods, these do not fully demonstrate the advantages of DarkOcc over existing illumination-robust occupancy prediction approaches. A comparison with relevant methods designed for nighttime or low-light occupancy prediction, such as LIAR (NeurIPS 2025), or a clearer discussion of their applicability to the indoor setting, would strengthen the experimental evaluation.

  4. Fragmented presentation of the methodology. The Method section is divided into too many small subsections, resulting in a fragmented presentation. Although the individual components are understandable, the connections between the motivation and the proposed designs are not sufficiently clear. A more coherent organization would help readers better understand how each component addresses the identified limitations.

Questions for authors

What evidence supports insufficient volumetric reasoning as the primary cause of occupancy omissions under low illumination? Could the authors provide comparisons with simpler voxel refinement methods or ablations without depth guidance and coarse-scale 3D contextual reasoning?

Since the voxel feature modulation in Eq. (3) depends on the base depth estimate, how sensitive is GOC to depth errors under severe low light? A depth-quality-stratified analysis or an experiment with perturbed/unconditioned depth would clarify whether the gains persist when the geometric prior itself is unreliable.

5 comments
AuthorsOpenReview ↗

Response to Reviewer #LowZ (1/5)

Thank you for recognizing the practical importance of indoor occupancy prediction across lighting conditions. We have carefully considered your comments and would like to address them as below.
We invite you to revisit the revised manuscript, where we have clarified the presentation throughout, redrawn the DarkOcc framework in Figure 2, and added a flowchart in Figure 3 showing how our team constructed and verified the real-world occupancy labels.


W1 [The performance trade-off between normal and low light and how MFA controls the strength of adaptation]

Multi-level feature adaptation (MFA) aims to improve low-light predictions while preserving accuracy under normal light. We measured how feature updates change with illumination on the same test frames and separately examined how gating and the update cap affect accuracy.

How much does MFA change the features under different lighting conditions? Table R1 reports the gate responses and update ratios at all four adaptation layers for the same Occ-ScanNet-mini test frames under four lighting conditions. The update ratio is the norm of the difference between the adapted and input features divided by the norm of the input features at that layer.

Table R1. Mean gate responses and mean feature update ratios at each layer under different lighting conditions. Layers 4, 11, 17, and 23 here (zero-based) correspond to layers 5, 12, 18, and 24 in Table 9 of the revised manuscript.

LightingMean local gateMean global gateLayer 4 updateLayer 11 updateLayer 17 updateLayer 23 update
Original0.470.150.54%0.55%0.21%0.04%
Dark10.590.441.48%1.50%0.62%0.12%
Dark20.730.862.68%2.68%1.26%0.24%
Dark30.850.992.99%2.99%1.60%0.30%

As the same frames become progressively darker, the mean update ratio increases at all four layers. At layers 4 and 11, the ratios rise from 0.54% and 0.55% under normal light to 2.99% at both layers under Dark3. The cap is triggered at a rate of 99.58% at both layers under Dark3. At layers 17 and 23, the mean updates under Dark3 are only 1.60% and 0.30%, respectively, and neither layer reaches the cap. These measurements show how gating adjusts the strength of the correction with illumination and how the 3% cap primarily constrains the layers with larger updates under severe low light.

How do gating and the update cap affect prediction accuracy? Using the illumination-adapted base (IAB), which includes full MFA and omits geometry-guided occupancy completion (GOC), as the reference, we separately remove illumination gating or the 3% cap (Table R2).

Table R2. Separate ablations of illumination gating and the feature update cap. Occupancy scores are reported as Geo-IoU / mIoU. Δ = variant − full MFA, in pp (percentage points).

Test conditionFull MFA (%)Without illumination gating (%)ΔWithout the 3% update cap (%)Δ
Occ-ScanNet-mini · Original61.47 / 56.6161.12 / 56.48−0.35 / −0.1361.29 / 56.51−0.18 / −0.10
Occ-ScanNet-mini · Dark359.02 / 52.9157.59 / 52.21−1.43 / −0.7058.53 / 52.72−0.49 / −0.19
Real indoor · normal light42.86 / 23.9341.61 / 23.46−1.25 / −0.4742.72 / 23.86−0.14 / −0.07
Real indoor · low light23.97 / 13.8721.22 / 12.76−2.75 / −1.1123.53 / 13.66−0.44 / −0.21

With the cap fixed at 3%, adding illumination gating raises Geo-IoU on paired frames by 0.35 pp under normal light and 1.43 pp under Dark3. The Geo-IoU gap between normal light and Dark3 therefore narrows from 3.53 to 2.45 pp while normal-light accuracy also improves. On the real data, the Geo-IoU / mIoU gains from gating are +2.75 / +1.11 pp under low light and +1.25 / +0.47 pp under normal light. The cap also improves both metrics under both lighting conditions on the synthetic and real data.


Please see the subsequent Authors' Rebuttal (2/5) for remaining responses.

AuthorsOpenReview ↗

Response to Reviewer #LowZ (2/5)

Continued: Authors’ Rebuttal (2/5)


W2 & Q1 & Q2 [Occupancy omissions under low light, the roles of predicted depth and 3D context, and GOC’s sensitivity to depth errors]

Thank you for raising this point. Figure 1(c) shows that even after feature adaptation, as the same test frames grow darker, the model misses more of the occupied voxels it correctly identified under normal light. You ask why these omissions arise and suggest ablations to examine the roles of depth guidance and 3D context in completion. We address these related concerns together with Q1 and Q2 below.

In the original submission, “volumetric reasoning” referred to reasoning about occupied structures and their spatial relationships throughout the 3D scene. Our wording could suggest that we attributed the omissions primarily to insufficient 3D reasoning. In the revision, we motivate GOC by the occupancy omissions that persist after feature adaptation and use ablations to examine how predicted depth and the surrounding 3D structure affect the final occupancy predictions.

Following your suggestion, we compared the full GOC with a generic 3D residual network of comparable capacity. In separate ablations, we removed the depth-aware feature modulation or the coarse-scale 3D branch. Table R3 reports each variant’s absolute scores and its differences from the full model.

(1) Contributions of predicted depth and 3D context to occupancy prediction

Table R3. Comparison with generic 3D refinement and ablations of GOC components on real indoor data. Occupancy scores are reported as Geo-IoU / mIoU. Δ = variant − full DarkOcc, in pp.

ConfigurationNormal light (%)Normal light ΔReal low light (%)Real low light Δ
Full DarkOcc46.38 / 25.53—29.12 / 15.56—
Capacity-matched generic 3D residual network in place of GOC46.14 / 25.42−0.24 / −0.1128.16 / 15.08−0.96 / −0.48
Remove only depth-aware feature modulation45.91 / 25.37−0.47 / −0.1628.59 / 15.39−0.53 / −0.17
Remove coarse branch, retain local branch and retrain45.88 / 25.35−0.50 / −0.1827.73 / 14.98−1.39 / −0.58

Capacity and generic 3D refinement. At comparable network capacity, the full GOC outperforms generic refinement under low light by 0.96 pp in Geo-IoU and 0.48 pp in mIoU. Under normal light, the corresponding gains are 0.24 and 0.11 pp. The advantage of GOC over generic refinement is therefore larger under low light.

Depth-aware modulation. Removing this modulation reduces Geo-IoU by 0.47 pp under normal light and 0.53 pp under low light. Depth-aware modulation therefore helps occupancy prediction under both lighting conditions.

Dual-scale context. The local branch operates at the original voxel resolution to preserve boundary detail. When a predicted structure is fragmented, the coarse-scale branch can draw on the surrounding 3D layout beyond a voxel’s immediate neighborhood. We removed the coarse-scale branch and retrained the model, retaining depth modulation, the local branch, and the fusion rule. Adding the coarse-scale branch improved Geo-IoU under both lighting conditions, with a larger gain under low light (+1.39 pp) than under normal light (+0.50 pp).


Please see the subsequent Authors' Rebuttal (3/5) for remaining responses.

AuthorsOpenReview ↗

Response to Reviewer #LowZ (3/5)

Continued: Authors’ Rebuttal (3/5)


(2) Does GOC retain its completion gains when the base depth estimates contain errors?

To address Q2, we used three experiments to examine how much of GOC’s completion gain remains when the base depth estimates contain errors.

(a) Removing explicit depth cues. We set all seven explicit depth cues supplied to GOC to zero (Table R4). On the paired test frames, the drop in Geo-IoU increases progressively from 0.11 pp under Original to 0.21 pp under Dark3. On the real low-light dataset, Geo-IoU and mIoU drop by 0.35 and 0.24 pp, respectively. The base model’s depth estimates thus provide useful cues even under severe low light.

Table R4. Occupancy accuracy after removing GOC’s explicit depth cues. Scores are reported as Geo-IoU / mIoU. Δ = score without explicit depth cues − full DarkOcc score, in pp.

Dataset and lightingFull DarkOcc (%)Without explicit depth cues (%)Δ
Occ-ScanNet-mini · Original64.20 / 57.3564.09 / 57.26−0.11 / −0.09
Occ-ScanNet-mini · Dark164.10 / 57.2663.94 / 57.17−0.16 / −0.09
Occ-ScanNet-mini · Dark263.67 / 56.4863.49 / 56.37−0.18 / −0.11
Occ-ScanNet-mini · Dark361.96 / 53.7361.75 / 53.59−0.21 / −0.14
Real indoor · normal light46.38 / 25.5346.15 / 25.48−0.23 / −0.05
Real indoor · low light29.12 / 15.5628.77 / 15.32−0.35 / −0.24

(b) Grouping frames by input depth error. We divided the 2,376 Dark3 test frames from Occ-ScanNet-mini into three equal groups by per-frame depth MAE. Within each group, we compared GOC enabled and disabled on the same 792 frames (Table R5). GOC improves accuracy in all three groups. Even in the group with the highest mean per-frame MAE (0.49 m), the gains are 3.56 pp in Geo-IoU and 1.11 pp in mIoU.

Table R5. GOC gains under Dark3, grouped by input depth error. The last column reports Geo-IoU / mIoU. Δ = accuracy with GOC enabled − accuracy with GOC disabled within the same group, in pp.

Depth error groupFramesMean frame MAE (m)Weighted MAE (m)Weighted AbsRelΔ Geo-IoU / Δ mIoU
Low error7920.360.360.24+2.11 / +0.67
Medium error7920.410.410.25+2.99 / +0.82
High error7920.490.490.26+3.56 / +1.11

Please see the subsequent Authors' Rebuttal (4/5) for remaining responses.

Reviewer_rrGQOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DarkOcc, an illumination-robust framework for indoor 3D occupancy prediction. DarkOcc introduces Multi-Level Feature Adaptation to improve visual representations under varying illumination conditions and Geometry-Guided Occupancy Completion to recover missing occupied regions through depth-aware voxel features and 3D contextual reasoning. Experiments on synthetic and real low-light datasets demonstrate improvements over existing methods.

Critical strengths
  1. This paper addresses a practically relevant problem of indoor occupancy prediction under low-light conditions. The observation that low illumination exacerbates occupancy incompleteness provides useful motivation.

  2. The experimental evaluation is relatively comprehensive, covering different illumination levels, monocular and embodied occupancy prediction.

Critical weaknesses
  1. The proposed modules are not sufficiently tailored to the unique challenges of low-light occupancy prediction. In particular, GOC mainly performs generic voxel refinement through depth-aware feature modulation and 3D contextual reasoning, without explicitly addressing illumination-induced geometric degradation. Table 4 shows that the performance improvements under normal illumination are nearly identical to those under severe low light

  2. The effectiveness of depth-aware feature modulation, dual-scale 3D contextual reasoning, and base-preserving residual fusion remains unclear. It is therefore difficult to determine whether the improvements arise from the proposed geometric guidance or simply from additional network capacity and 3D refinement.

Questions for authors
  1. Could the authors provide more detailed ablations of GOC, particularly comparisons with a generic 3D refinement network of comparable capacity and variants without depth-aware feature modulation?
4 comments
AuthorsOpenReview ↗

Response to Reviewer #rrGQ (1/2)

Thank you for recognizing the importance of our observation that more occupied voxels are missed as illumination decreases. We appreciate the opportunity to explain how this observation motivated GOC and to quantify its gains under different lighting conditions.
We invite you to revisit the revised manuscript, where we have clarified the presentation throughout, redrawn the DarkOcc framework in Figure 2, and added a flowchart in Figure 3 showing how our team constructed and verified the real-world occupancy labels.


W1 [How GOC addresses occupancy omissions in low light and how its gains vary with lighting]

The paired experiment in Figure 1(c) motivated geometry-guided occupancy completion (GOC). We evaluated the same Occ-ScanNet-mini test frames at progressively lower light levels. Even after multi-level feature adaptation (MFA), as the inputs grew darker, more voxels correctly identified as occupied in normal light were misclassified as free. We therefore designed GOC to revisit voxels that the base model predicts as free and use predicted depth and the surrounding 3D structure to determine which should be recovered as occupied. The GOC ablation also revealed an unexpected benefit. The base model misses occupied voxels even in normal light, so the same mechanism can improve predictions there as well. This benefit is consistent with our aim of improving robustness to low light while preserving accuracy in normal light. Table R1 reports accuracy on real indoor data before and after adding GOC.

Table R1. Occupancy prediction accuracy on real indoor data before and after adding GOC. Metrics are reported as Geo-IoU / mIoU. IAB is the illumination-adapted base model. Δ = DarkOcc − IAB, in pp (percentage points).

Dataset and lightingIAB (%)DarkOcc (%)Gain from GOC (pp)
Real indoor · normal light42.86 / 23.9346.38 / 25.53+3.52 / +1.60
Real indoor · low light23.97 / 13.8729.12 / 15.56+5.15 / +1.69

As Table 5 of the revised manuscript shows, adding GOC to the base model yielded a slightly larger Geo-IoU gain under Dark3 than under normal light on Occ-ScanNet-mini. The difference was more pronounced on the real indoor test set (Table 3), where GOC improved Geo-IoU by 3.52 pp under normal light and 5.15 pp under low light. Below, we compare the full GOC module with a capacity-matched 3D refinement network and separately ablate GOC’s coarse-scale branch to quantify its contribution under both normal and low light.


Please see the subsequent Authors' Rebuttal (2/2) for remaining responses.

AuthorsOpenReview ↗

Response to Reviewer #rrGQ (2/2)

Continued: Authors’ Rebuttal (2/2)


W2 & Q1 [Capacity-matched comparison with generic 3D refinement and contributions of the GOC components]

Thank you for suggesting a capacity-matched comparison and ablations of the individual components.

Table R2 reports the four ablations. A1 directly compares full GOC with a generic 3D residual network of comparable capacity. A2 isolates the contribution of depth-aware feature modulation. A3 and A4 test the contributions of coarse-scale context and the update rule that preserves the base model’s occupied predictions, respectively.

Table R2. GOC component ablations on real indoor data. Metrics are reported as Geo-IoU / mIoU. Δ = variant − full DarkOcc, in pp.

MethodNormal light (%)Normal light ΔReal low light (%)Real low light Δ
DarkOcc46.38 / 25.53—29.12 / 15.56—
A1 Capacity-matched generic 3D residual network in place of GOC46.14 / 25.42−0.24 / −0.1128.16 / 15.08−0.96 / −0.48
A2 Remove only depth-aware feature modulation45.91 / 25.37−0.47 / −0.1628.59 / 15.39−0.53 / −0.17
A3 Remove coarse branch, retain local branch45.88 / 25.35−0.50 / −0.1827.73 / 14.98−1.39 / −0.58
A4 Update and supervise all voxels with valid projections46.29 / 25.49−0.09 / −0.0429.00 / 15.51−0.12 / −0.05

Capacity and generic 3D refinement (A1). Under low light, full GOC outperformed generic refinement at comparable capacity by 0.96 pp in Geo-IoU and 0.48 pp in mIoU. The corresponding margins under normal light were smaller, at 0.24 and 0.11 pp. The advantage over generic refinement was therefore larger under low light.

Depth-aware modulation (A2). Removing this modulation reduced Geo-IoU by 0.47 pp under normal light and by 0.53 pp under low light, indicating that depth-aware modulation improves occupancy prediction under both lighting conditions.

Dual-scale context (A3). The local branch operates at the original voxel resolution to preserve boundary detail. When a predicted structure is fragmented, the coarse-scale branch can draw on the surrounding 3D layout beyond a voxel’s immediate neighborhood. We removed the coarse-scale branch and retrained the model, retaining depth modulation, the local branch, and the fusion rule. Adding the coarse-scale branch improved Geo-IoU under both lighting conditions, with a larger gain under low light (+1.39 pp) than under normal light (+0.50 pp).

Residual fusion that preserves the base model’s occupied predictions (A4). We extended both the update region and the corresponding supervision to all voxels with a valid projection, while keeping the architecture, parameter count, and residual addition unchanged. The two settings achieved similar accuracy. Retaining the mask improved Geo-IoU by 0.09 and 0.12 pp under normal light and low light, respectively. We retain this constraint primarily because it focuses GOC updates on regions that may need completion while achieving comparable accuracy.


We hope these results address your questions and comments. We would be happy to discuss them further during the rebuttal period.

I thank the authors for their detailed response. The clarifications are reasonable. I am inclined to maintain my original score

AuthorsOpenReview ↗

Thank you for your follow-up

Thank you for reading our response and for following up so promptly.

We are pleased to hear your positive assessment of our clarifications. Your questions helped us clarify why we introduced GOC. One unexpected finding was that GOC also recovers missed occupied regions under normal lighting. On our real-world indoor dataset, we now show more clearly how much GOC improves occupancy prediction under low light and how much its coarse-scale 3D branch contributes to that gain. Thank you for your valuable feedback.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Reviewers, AC, SAC, and PC,

Thank you for reviewing and handling our paper. We summarize the main findings below and point to the corresponding experiments in the revised manuscript and our individual responses.


Paper Overview and Strengths

[Paper Overview] DarkOcc predicts indoor 3D semantic occupancy from RGB under normal and low light. Multi-Level Feature Adaptation (MFA) adjusts encoder features to the illumination. Geometry-Guided Occupancy Completion (GOC) uses predicted depth and 3D context to recover occupied regions still missed after adaptation.

[Strengths] All three reviewers recognized the practical importance of the task. rrGQ and LowZ highlighted the finding that more occupied voxels are missed as the same scenes grow darker. rrGQ also noted the breadth of the monocular and embodied evaluations, and agiB highlighted the value of the real low-light dataset for robotics.


Additional Experiments Addressing Reviewers’ Concerns

  • [Matched training conditions] (agiB). We retrained EmbodiedOcc, SplatSSC, and FLM-Occ under the same mixed-light training conditions as DarkOcc. On real low-light data, DarkOcc outperformed all three and mixed-light GPOcc in both metrics, exceeding the best baseline Geo-IoU by 9.40 percentage points (pp) (Tables 1 and 3).
  • [Illumination gating] (LowZ, agiB). Adding the illumination gate improved Geo-IoU on real indoor data by 2.75 pp under low light and 1.25 pp under normal light (LowZ, Table R2). The larger low-light gain suggests that gating helps MFA adapt visual features more effectively to dimly lit scenes.
  • [Geometry-guided completion] (rrGQ, LowZ, agiB). On real low-light data, GOC outperformed capacity-matched 3D refinement by 0.96 pp in Geo-IoU. Removing depth modulation or coarse-scale context reduced Geo-IoU by 0.53 pp and 1.39 pp, respectively (rrGQ, Table R2).
  • [LIAR transfer] (LowZ, agiB). Training GPOcc with LIAR’s SLLIE and 2D-IGS improved Dark3 Geo-IoU by 6.84 pp. DarkOcc scored higher than both tested LIAR configurations in Geo-IoU and semantic mIoU at all four light levels (Tables 1 and 19).

Summary of Responses to Reviewers’ Questions

  • [What motivated GOC?] (rrGQ, LowZ, agiB). As the same frames darken, the base model still misses more occupied voxels after MFA (Figure 1(c)), motivating explicit occupancy completion with GOC. On real indoor data, adding GOC after MFA improved Geo-IoU by 5.15 pp under low light. An unexpected benefit was the 3.52 pp gain under normal light, consistent with our goal of improving low-light robustness while preserving normal-light accuracy (Table 3). Removing the coarse-scale branch reduced Geo-IoU more under low light (1.39 pp) than under normal light (0.50 pp) (rrGQ, Table R2).
  • [Sensitivity to depth errors] (LowZ). On real low-light data, adding explicit depth cues improved Geo-IoU by 0.35 pp (LowZ, Table R4). Across the low-, medium-, and high-depth-error groups on Dark3, enabling GOC increased Geo-IoU by 2.11, 2.99, and 3.56 pp, respectively (Table 15). With upstream predictions fixed, GOC gained 2.94 pp in Geo-IoU with unperturbed depth and 2.89 pp with either 0.8× depth scaling or the strongest tested noise (σ = 0.32 m) (Appendix C.3, Table 16).

Revisions During Rebuttal

We revised the text throughout the manuscript for clarity. We now clarify that darker inputs led to more occupancy omissions even after MFA, motivating GOC.

  • [Framework diagram] We revised the writing throughout and redrew Figure 2. Section 3 now follows one RGB frame through the model, with separate panels showing feature adaptation and occupancy completion.
  • [Real indoor dataset] The new flowchart (Section 4.1, Figure 3) shows how we reconstructed ten indoor scenes from 19 RGB-D recordings, labeled them across views, and produced occupancy labels for 8,208 frames. Several team members annotated and reviewed the geometry and semantic labels over multiple rounds, correcting discrepancies against the recordings (Appendix D.1). Figure 7 shows the reviewed references for all ten scenes.
  • [Ablation studies] Section 4.6 adds real-data ablations of illumination gating, depth-aware modulation, and coarse 3D context (Table 6) alongside the cumulative ablation (Table 5). Appendix A.2 reports gate responses and feature update magnitudes (Table 9).

We provide code and sample data from our real-world indoor occupancy dataset in an anonymous repository (https://anonymous.4open.science/r/DarkOcc-review-3624/) for review. We plan to release the full codebase and dataset upon publication to support further research on indoor occupancy prediction under low light.

We hope our responses and revisions address the reviewers’ concerns. If any questions remain, please let us know and we will respond promptly. All changes are highlighted in blue in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.