- 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
-
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.
-
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.
-
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
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 condition | No light gate (%) | Full MFA (%) | Gain from gating (pp) |
|---|---|---|---|
| Occ-ScanNet-mini · Original | 61.12 / 56.48 | 61.47 / 56.61 | +0.35 / +0.13 |
| Occ-ScanNet-mini · Dark3 | 57.59 / 52.21 | 59.02 / 52.91 | +1.43 / +0.70 |
| Real indoor · normal light | 41.61 / 23.46 | 42.86 / 23.93 | +1.25 / +0.47 |
| Real indoor · low light | 21.22 / 12.76 | 23.97 / 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.
| Configuration | Normal light (%) | Normal light gain (pp) | Real low light (%) | Real low light gain (pp) |
|---|---|---|---|---|
| Full GOC | 46.38 / 25.53 | — | 29.12 / 15.56 | — |
| Capacity-matched generic 3D residual network in place of GOC | 46.14 / 25.42 | +0.24 / +0.11 | 28.16 / 15.08 | +0.96 / +0.48 |
| Remove coarse branch, retain local branch | 45.88 / 25.35 | +0.50 / +0.18 | 27.73 / 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.
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.
| Method | Year-Venue | Outdoor occupancy prediction | Indoor occupancy prediction | Single-view RGB input | Illumination gating at multiple encoder stages | Geometry-Guided Occupancy Completion (GOC) |
|---|---|---|---|---|---|---|
| LIAR | 2025-NeurIPS | ✓ | — | — | — | — |
| DepthDark | 2025-ACM MM | — | — | ✓ | — | — |
| NightFormer | 2024-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.
| Method | Year-Venue | Original | Dark1 | Dark2 | Dark3 |
|---|---|---|---|---|---|
| GPOcc | 2026-CVPR | 63.16 / 55.85 | 60.71 / 53.63 | 52.84 / 45.61 | 35.65 / 28.58 |
| GPOcc + LIAR (SLLIE) | 2025-NeurIPS | 62.98 / 55.62 | 60.23 / 53.06 | 51.98 / 44.81 | 35.29 / 28.40 |
| GPOcc + LIAR (SLLIE + 2D-IGS, trained) | 2025-NeurIPS | 63.48 / 56.07 | 61.58 / 54.26 | 55.74 / 47.59 | 42.49 / 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.
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 (%).
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.