- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
This paper proposes OWL-Net for low-light open-vocabulary segmentation, which uses RAW data only during training while relying solely on RGB input at inference. It consists of a REC module that encodes RAW data by decomposing and reweighting it at the bit-plane level, and an RRE module that aligns low-light RGB features to the RAW features via a cosine loss. Training is performed on low-light RGB/RAW pairs synthesized from COCO-Stuff. The authors also construct LOVS, a dataset of 1,000 aligned normal-/low-light image pairs (220 classes) captured with an ND filter, and report consistent improvements on synthetic benchmarks and LOVS when OWL-Net is applied to three OVS frameworks.
Critical strengths
S1. Practical problem setting. Low-light OVS is a largely unexplored area, and the design of using RAW only for training while performing RGB-only inference keeps the deployment cost low (inference time 126→133 ms in Table 5).
S2. LOVS dataset. No existing real-world low-light OVS benchmark provides pixel-aligned normal-/low-light pairs together with RAW data and masks for 220 classes. If released, it would be a lasting contribution to the community independent of the proposed method.
S3. Consistent plug-and-play gains. OWL-Net yields improvements in the same direction across all benchmarks on three structurally different frameworks (CAT-Seg, OVRCOAT, OpenWorldSAM).
Critical weaknesses
W1. It is not verified that the gains come from RAW. The training RAW data are synthesized by inverting 8-bit COCO sRGB images through an inverse ISP (Appendix K), so they cannot contain more information than the original normal-light images. Thus, the benefit of the RAW branch during training may not be "the fine-grained information captured by the sensor" (L48–53), but merely a clean signal before noise injection and finer quantization. Moreover, training on synthesized low-light data naturally provides paired normal-light images, and using them as a teacher is a more direct alternative than RAW (cf. DAI-Net [1]). Nevertheless, neither such a comparison nor an ablation that trains the same RRE+fusion architecture without the RAW alignment loss is provided. The facts that feeding real RAW at inference yields only +0.3 mIoU, and that replacing RRE with plain convolutions brings mIoU back to the baseline level (Table 3, 48.1), also suggest that the gains may stem from the capacity of the added modules rather than from RAW supervision.
W2. Inappropriate choice of baselines. The baselines in Table 1 are only generic OVS models that do not account for low light. Low-light adaptation methods (e.g., NightAdapter [2], FeatEnHancer [3]) and RAW-based methods [4, 5], which the authors themselves cite, are not compared. The two-stage comparison (Table 5) includes only two enhancers, both of which achieve higher mIoU than OWL-Net (51.0 and 52.9 vs. 49.2), while lightweight enhancers (e.g., SCI [6]) and RAW-domain restoration are missing. The "best" results also largely come from the strength of OpenWorldSAM itself, and without baselines matched in added modules and training iterations, fairness is hard to judge. Furthermore, the real-world low-light datasets the authors list in Fig. 3 (ACDC-Night, Dark Zurich, NightCity, LIS) are not used for evaluation.
W3. Small gains and unclear evaluation reliability. On LOVS, the improvement over OpenWorldSAM is +1.1 mIoU. The seemingly larger hIoU gain (+5.3) is highly variable because there are only 14 seen classes. There are no repeated runs with different seeds, and LOVS contains only 1,000 images with a long-tailed distribution, making it difficult to judge the significance of such differences.
W4. Realism of LOVS. An ND filter only attenuates illumination uniformly, so LOVS lacks characteristics of real night scenes such as non-uniform illumination, light sources and glare, and color casts from artificial lighting. It is captured with a single camera and a single attenuation level (128×), limiting the scope of generalization. The gap with the attenuation used for training synthesis (~20×) and possible degradation under normal light are not analyzed.
Minor. The increasing cosine similarity in Appendix G only shows that the optimized loss decreases. The PSNR of 34 dB measured between nearly black images (Table 8) is weak evidence of synthesis fidelity.
Questions for authors
- Could you report results of training RRE+fusion without the RAW alignment loss, and of using the paired normal-light RGB as the teacher? It would be helpful to show that the RAW teacher is meaningfully better.
- Do you have results comparing REC with feeding float RAW and with exposure amplification followed by a gamma/log transform?
- Could you provide results of attaching low-light adaptation methods (e.g., NightAdapter [2], FeatEnHancer [3], DAI-Net [1]) to the same frameworks, of RAW-based methods and RAW-domain restoration + OVS, and on existing real-world low-light datasets?
If these questions and the main weaknesses are addressed, I will raise my score.
References
[1] Z. Du et al., "Boosting Object Detection with Zero-Shot Day-Night Domain Adaptation," CVPR 2024.
[2] Q. Bi et al., "NightAdapter: Learning a Frequency Adapter for Generalizable Night-time Scene Segmentation," CVPR 2025.
[3] K. A. Hashmi et al., "FeatEnHancer: Enhancing Hierarchical Features for Object Detection and Beyond Under Low-Light Vision," ICCV 2023.
[4] Z.-Y. Li et al., "Towards RAW Object Detection in Diverse Conditions," CVPR 2025.
[5] L. Chen et al., "Instance Segmentation in the Dark," IJCV 2023.
[6] L. Ma et al., "Toward Fast, Flexible, and Robust Low-Light Image Enhancement," CVPR 2022.
9 comments
Response Part 2
Q1.2: The effectiveness of RAW feature alignment.
To investigate whether the improvements mainly arise from the additional network components, we further conduct an ablation experiment by removing the cosine feature alignment loss while retaining the REC, RRE, and dark-aware fusion modules. In this setting, the RGB and RAW branches are still optimized through their respective segmentation losses, but no explicit feature alignment is imposed between them. This comparison allows us to isolate the contribution of RAW-to-RGB feature alignment from that of the additional network architecture. The experimental results on LOVS are summarized below.
| Method | mIoU ↑ | hIoU ↑ |
|---|---|---|
| OpenWorldSAM | 48.1 | 46.0 |
| OWL-Net (w/o Alignment Loss) | 48.4 | 48.9 |
| OWL-Net (Ours) | 49.2 | 51.3 |
As shown in the table, removing the feature alignment loss leads to performance degradation of 0.8 mIoU and 2.4 hIoU compared with the full OWL-Net. These results demonstrate that the performance gains are not solely attributable to the additional network components, but also benefit from explicit RAW-to-RGB feature alignment, which enables more effective transfer of RAW-domain information to the enhanced RGB representations.
Q1.3: Clarification of the RRE effectiveness.
Regarding the relatively small improvement (+0.3 mIoU) obtained by introducing real RAW input during inference, we would like to clarify that this experiment is designed to evaluate how closely the RRE-enhanced RGB representations approximate the guidance provided by real RAW features. As reported in Table 3 in the original manuscript, OWL-Net achieves 49.2 mIoU and 51.3 hIoU using only RGB input, while additionally incorporating real RAW input during inference achieves 49.5 mIoU and 51.9 hIoU. The relatively small performance gap suggests that RRE module effectively transfers useful RAW-domain guidance into the RGB representations during training, thereby reducing the reliance on additional RAW inputs during inference.
Meanwhile, the comparison with plain convolutions aims to verify that the performance improvements arise from the tailored design of RRE, rather than simply increased model parameter. As shown in Table 3 of the original submission (now Table 5 in the revised manuscript), replacing RRE with plain convolution yields only 48.1 mIoU and 48.8 hIoU. Even when using parameter-matched convolutional layers, the performance remains at 48.8 mIoU and 50.8 hIoU, below the full RRE module (49.2 mIoU and 51.3 hIoU). These comparisons support the benefit of the proposed local-to-global representation enhancement design over simple convolutional alternatives.
We sincerely appreciate the reviewer's constructive comments, which help us clarify the contributions of RAW-based supervision and the proposed RRE module.
Response Part 3
Q2: The fairness of baseline comparisons and evaluation protocols
Response: Thank you for your comments. We believe that some aspects of our experimental settings and comparison protocols may have been misunderstood. We would like to provide the following clarifications.
Q2.1: Comparisons with existing low-light perception methods.
First, we would like to emphasize that our work focuses on low-light open-vocabulary segmentation, which differs fundamentally from the tasks considered in the suggested methods. Specifically, Du et al. [1] investigate zero-shot day-to-night domain adaptation for object detection, NightAdapter [2] focuses on nighttime semantic segmentation, FeatEnHancer [3] addresses low-light perception tasks such as object detection and semantic segmentation, while [4] and [5] investigate RAW-based object detection and instance segmentation, respectively. None of these methods is originally designed for open-vocabulary segmentation, where the model needs to recognize arbitrary text-defined categories, including unseen categories. Therefore, their original experimental results are not directly comparable with ours under the same evaluation protocol.
More importantly, our comparisons in Table 1 are not conducted by directly applying models pretrained on normal-light images to low-light data. Instead, all compared methods are initialized from their respective pretrained models and further fine-tuned using the same synthetic low-light COCO-Stuff training data. For each baseline and its OWL-Net counterpart, we maintain identical training configurations, including the optimizer, learning-rate schedule, batch size, and training iterations. Therefore, our evaluation provides a fair comparison under consistent low-light training and open-vocabulary segmentation settings.
Response Part 1
Dear Reviewer ZwzA,
We sincerely thank you for your thoughtful and constructive review. We particularly appreciate your recognition of the practical value of our research problem and the consistent improvements achieved by OWL-Net across different frameworks. Your constructive suggestions are so valuable to us. We are honored to have the opportunity to clarify the points you have raised. Below is our detailed response to answer your concerns.
Q1: The effectiveness of RAW supervision and RRE module
Response: Thank you for your comments. According to your suggestion, we provide additional experiments and clarifications to address these concerns.
Q1.1: The effectiveness of RAW supervision.
We would like to clarify that our RAW-assisted learning strategy aims to exploit the fine-grained information preserved in RAW representations, rather than introduce additional information beyond the original images. Compared with the corresponding rendered low-light RGB images, the synthetic RAW data retain fine-grained signal variations that may be weakened during subsequent image processing. Therefore, under extremely low-light condition, RAW data could preserve richer structural details that may be weakened in the degraded RGB images. Moreover, as illustrated in Fig. 5 of the Appendix, RAW data preserve discernible object structures across a wider range of bit planes (particularly bits 2–8), whereas the corresponding low-light RGB images retain visible structural information mainly in the lowest bit planes (bits 0–2). These observations further support the effectiveness of exploiting RAW-domain information to guide RGB representation learning under challenging low-light conditions.
To further verify this, following the reviewer's suggestion, we introduce an additional experiment by replacing the synthetic RAW input to REC with the corresponding original normal-light RGB image. The remaining training framework and feature alignment strategy are kept unchanged, allowing us to compare the effectiveness of normal-light RGB supervision and RAW supervision.
As shown in the table, OWL-Net with RAW supervision outperforms the normal-light RGB teacher by 0.6 mIoU and 1.2 hIoU, demonstrating the effectiveness of RAW-based guidance. This may be attributed to the considerable gap between normal-light and low-light RGB images. Although normal-light RGB images contain clearer visual information, their representations exhibit a considerable gap from those of low-light RGB images. In contrast, low-light RAW data preserve informative structural details while maintaining greater consistency with the target low-light conditions, thereby providing more effective guidance for RGB representation learning.
Thanks for your suggestion. We have added this experiment in the revised paper. (Presented as Table 3 in the main paper)