acceptodds
Under review as a conference paper at ICLR 2027

Open-Vocabulary Segmentation in the Dark

10 views

Abstract

Open-vocabulary segmentation has achieved substantial progress in recent years, yet existing methods are mainly developed for well-illuminated images and suffer significant performance degradation under low-light conditions. In this work, we propose OWL-Net, a RAW-assisted framework for open-vocabulary segmentation in the dark. OWL-Net exploits the fine-grained intensity and structural information preserved in RAW data during training, while requiring only RGB input during inference. Specifically, we introduce a RAW EnCoding (REC) module that represents large-range RAW values as the binary bit sequence and extracts RAW features directly at the bit level, enabling critical details within RAW data to be better preserved during feature extraction. Meanwhile, to avoid the reliance on RAW input during inference, we further develop an RGB Representation Enhancement (RRE) module to progressively recover the overwhelmed details and lift the degraded RGB features to RAW-aware space. In addition, we construct LOVS, a real-world Low-light Open-Vocabulary Segmentation dataset containing 1,000 aligned normal-/low-light image pairs, covering 220 semantic categories with 6,949 object-level masks. Extensive experiments on both synthetic low-light benchmarks and LOVS demonstrate that OWL-Net consistently improves different open-vocabulary segmentation frameworks under low-light conditions, without requiring RAW input or explicit image enhancement during inference. Our code and dataset will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Low-Light Perception, Segmentation in the Dark

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ZwzAOpenReview ↗
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
  1. 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.
  2. Do you have results comparing REC with feeding float RAW and with exposure amplification followed by a gamma/log transform?
  3. 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
AuthorsOpenReview ↗

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.

MethodmIoU ↑hIoU ↑
OpenWorldSAM48.146.0
OWL-Net (Normal-light RGB Teacher)48.650.1
OWL-Net (Ours)49.251.3

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)

AuthorsOpenReview ↗

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.

MethodmIoU ↑hIoU ↑
OpenWorldSAM48.146.0
OWL-Net (w/o Alignment Loss)48.448.9
OWL-Net (Ours)49.251.3

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.

AuthorsOpenReview ↗

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.

Reviewer_9ShcOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

OWL-Net uses RAW supervision during training to improve open-vocabulary segmentation of low-light RGB images. It introduces a bit-plane RAW encoder, an RGB feature enhancement module, and LOVS, a paired benchmark collected through controlled low-light capture.

Critical strengths
  • The paper addresses a useful extension of open-vocabulary segmentation. The paired images and broad category coverage make LOVS a potentially valuable benchmark.
  • OWL-Net improves several existing segmentation frameworks while retaining RGB-only inference. The component studies and comparisons with enhancement-then-segmentation pipelines provide useful evidence of its practical value.
Critical weaknesses
  • The evaluation mainly uses synthetically darkened versions of standard OVS datasets. LOVS provides real sensor captures, but its single-camera ND-filter protocol leaves generalization to naturally occurring low-light conditions unclear. Results across illumination and noise levels would help establish what the “in the dark” setting actually covers.
  • The source of the improvement is not sufficiently isolated. A control retaining the full RGB architecture without RAW supervision is needed. A teacher using the original normal-light RGB images would also be informative, especially since the training RAW is synthesized from those images.
  • Appendix E.3 describes the original encoders and mask decoder as frozen, while also stating that the baselines are retrained. The trainable modules and starting checkpoints for each baseline and its OWL-Net counterpart need to be specified more clearly.
  • The RGB inputs in Figure 4 are almost invisible at normal display settings, making the remaining scene structure difficult to assess. I suggest adding clearly marked brightened views for inspection and specifying the actual input value range and preprocessing.
  • REENet (Huang et al., TIP 2022) already uses RAW guidance during training with RGB-only inference, but is missing from the discussion.
Questions for authors
  • How does performance vary with illumination and noise severity, and is there evidence that the gains extend to natural low-light scenes beyond the LOVS capture protocol?
  • How much improvement remains when the full RGB architecture is trained without RAW supervision, or when the teacher uses the original normal-light RGB images?
5 comments
AuthorsOpenReview ↗

Response Part 1

Dear Reviewer 9Shc,

Thank you for your valuable comments. We appreciate the chance to clarify and address your concerns, which we believe will enhance our paper. Below, we respond in detail to each point you have raised:

Q1: The collection of our LOVS dataset

Response: Thank you for your valuable comment. We would like to clarify that our controlled acquisition strategy is primarily designed to ensure reliable pixel-level annotations. Specifically, annotating objects directly in extremely dark images is challenging, particularly for fine-grained categories and object boundaries. Therefore, we capture normal-light images under sufficient illumination and then use an ND filter to obtain spatially aligned low-light counterparts. This allows us to obtain accurate segmentation masks from normal-light images and directly transfer them to the corresponding low-light images. Notably, similar ND-filter-based acquisition strategies are also adopted in recent studies, including [R1] and [R2], to construct paired normal-light and low-light data.

To verify the generalization across different illumination and noise conditions, following the reviewer's suggestion, we conduct additional experiments using images captured with a 6-stop ND filter (64× attenuation), which are also retained during the construction of LOVS. Compared with the original 7-stop setting (128× attenuation), these images provide a different illumination level and noise conditions, allowing us to further examine the robustness of OWL-Net under varying low-light degradation. The results are summarized below.

ND SettingMethodmIoU ↑hIoU ↑
ND6 (64×)OpenWorldSAM52.853.7
OWL-Net54.055.9
ND7 (128×)OpenWorldSAM48.146.0
OWL-Net49.251.3

As shown in the table, OWL-Net consistently outperforms OpenWorldSAM under both attenuation settings, achieving mIoU improvements of 1.2 and 1.1 under ND6 and ND7, respectively. These results demonstrate the robustness of OWL-Net across illumination and noise levels. Thanks for your valuable comment. We have added these results and the corresponding analysis to Appendix L (Table 12) of the revised manuscript.

We appreciate the reviewer's constructive suggestion. We will further expand LOVS by incorporating diverse cameras, exposure settings, and natural nighttime scenes to support more comprehensive evaluations in future work.

[R1] CVPR2024-Towards Robust Event-guided Low-Light Image Enhancement: A Large-Scale Real-World Event-Image Dataset and Novel Approach

[R2] ICCV2025-Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring

AuthorsOpenReview ↗

Response Part 2

Q2: The effectiveness of REC module

Response: Thank you for your insightful comment. To address this concern, we conduct two additional experiments. (1) Directly ablate REC module: We remove the REC module and its associated supervision while retaining the complete RGB processing architecture, including the RRE module and dark-aware spatial fusion module. The remaining network is trained using only the segmentation objective. This experiment helps determine whether the improvements primarily arise from the additional network components. (2) Normal-light RGB supervision: We replace the synthetic RAW input to REC with the corresponding original normal-light RGB image, while retaining the same feature alignment strategy. In this setting, the normal-light RGB serves as the teacher to guide the feature learning of low-light RGB image. The results on LOVS are summarized below.

MethodmIoU ↑hIoU ↑
OpenWorldSAM48.146.0
OWL-Net (ablate REC module)48.449.2
OWL-Net (Normal-light RGB Teacher)48.650.1
OWL-Net (RAW Teacher, Ours)49.251.3

As shown in the table, removing RAW supervision or replacing it with normal-light RGB guidance leads to performance degradation, demonstrating the effectiveness of RAW-based supervision. In particular, using normal-light RGB images as input to the REC module yields lower performance than our RAW-guided approach. This may be attributed to the representation 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.

We sincerely appreciate the reviewer's constructive suggestion, which helps strengthen the experimental analysis and clarify the source of the performance improvements. According to your suggestion, we have added the experimental results in our revised paper. (Presented as Table 3 in the main paper)

Q3: Clarification of the fine-tuning configurations

Response: Thanks for your question. For each framework, the baseline and its OWL-Net counterpart are initialized from the same official pretrained checkpoint, which are downloaded from the corresponding official GitHub repositories. During the training process, rather than fine-tuning all model parameters, we follow the original training configurations of each framework and update only the designated trainable modules.

Specifically, for each baseline and its OWL-Net counterpart, the original trainable modules remain identical. The only difference is that OWL-Net additionally introduces three trainable modules: REC, RRE, and dark-aware spatial fusion. The detailed configurations are summarized below.

FrameworkTrainable modules in both baseline and OWL-NetAdditional trainable modules in OWL-Net
CAT-SegCost aggregation; Upsampling guidanceREC; RRE; Dark-aware fusion
OVRCOATMask generator (Pixel decoder; Mask Transformer; Prediction heads)REC; RRE; Dark-aware fusion
OpenWorldSAMLanguage adapter (MLP projector; Soft prompting Transformer)REC; RRE; Dark-aware fusion

All other parameters, including the pretrained visual and text encoders, remain frozen. Moreover, each baseline and its OWL-Net counterpart share identical training settings, including the optimizer, learning-rate schedule, batch size, and number of training iterations, ensuring a fair comparison. According to your valuable suggestion, we have further introduced these fine-tuning details in Table 7 of the Appendix E.3.

AuthorsOpenReview ↗

Response Part 3

Q4: The visualization issue

Response: Thank you for your valuable suggestion. We would like to clarify that the RGB inputs in Fig. 4 are intentionally displayed without brightness enhancement to faithfully reflect the challenging low-light conditions considered in our work. Although the visual details are difficult to distinguish in these extremely dark images, the ground-truth segmentation masks are obtained from their spatially aligned normal-light counterparts, where the object structures and boundaries are clearly visible. Therefore, the ground-truth visualizations in Fig. 4 provide clear references for assessing the segmentation results.

Meanwhile, this further highlights the rationale behind our ND-filter-based acquisition strategy (discussed in Q1). Since our study focuses on open-vocabulary segmentation under extremely challenging low-light conditions, obtaining spatially aligned normal-light references is essential for ensuring reliable pixel-level annotations. Therefore, we adopt a controlled ND-filter-based acquisition strategy to collect severely degraded low-light images while preserving accurate annotations through their normal-light counterparts.

Thanks for your valuable comment. We have revised the Fig. 4 in the manuscript according to your suggestion.

Q5: Comparison with REENet

Response: Thank you for pointing out this relevant work. We acknowledge that REENet [R3] also explores the use of RAW guidance during training while maintaining RGB-only inference. However, REENet and OWL-Net address fundamentally different downstream tasks and exploit RAW information for different purposes.

Specifically, REENet focuses on low-light image enhancement. During training, it exploits paired RAW data and exposure-related information to guide the Unprocess-Enhance-Process pipeline, learning to transform low-light sRGB images into well-exposed ones through linear-domain enhancement. Specifically, in this process, RAW data is utlized to provide supervision for the Unprocess network to map low-light sRGB images into the linear RGB domain. Once the training is done, REENet takes only low-light sRGB images as input to perform image enhancement.

In contrast, OWL-Net addresses a different task, i.e., low-light open-vocabulary segmentation, and exploits RAW information to guide the semantic representation learning rather than image reconstruction. Specifically, our REC module decomposes high-bit-depth RAW data into binary bit planes and adaptively encodes their significance across different Bayer channels, extracting informative RAW-domain representations. Moreover, our RRE module progressively enhances degraded RGB features through local and global feature compensation, with cosine feature alignment to the corresponding RAW representations. Through this process, RAW-domain knowledge is transferred into RGB representations to facilitate more reliable visual-semantic correspondence. Importantly, unlike REENet, which explicitly reconstructs enhanced images, OWL-Net directly performs open-vocabulary segmentation without an explicit image-enhancement stage during inference, relying solely on the low-light RGB input.

Since REENet is designed and evaluated for image enhancement rather than open-vocabulary segmentation, a direct quantitative comparison is not appropriate. Moreover, the original paper does not provide publicly available source code, making faithful reproduction and adaptation to our OVS framework difficult.

We appreciate the reviewer for bringing this relevant work to our attention. We have included REENet in the Related Work for further discussion.

[R3] TIP2022-Towards Low Light Enhancement With RAW Images

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

This paper investigates open-vocabulary segmentation under challenging low-light conditions and proposes OWL-Net, a RAW-assisted framework that requires only RGB input during inference. Specifically, it introduces a RAW EnCoding (REC) module to preserve informative bit-level details and an RGB Representation Enhancement (RRE) module to transfer RAW-aware knowledge into RGB features. The authors also construct LOVS, a real-world low-light segmentation dataset containing 220 categories. Experiments on synthetic and real-world benchmarks demonstrate consistent improvements across different open-vocabulary segmentation frameworks.

Critical strengths
  1. Practical motivation and valuable dataset contribution. The paper addresses an important yet relatively underexplored problem in open-vocabulary segmentation. The newly constructed LOVS dataset, containing 1,000 image pairs and 220 semantic categories, provides a useful benchmark for evaluating segmentation under extreme low-light conditions.

  2. Effective framework with comprehensive evaluation. OWL-Net effectively leverages RAW information during training while maintaining RGB-only inference. The consistent improvements across multiple segmentation frameworks and benchmarks demonstrate the practical value and generality of the proposed approach.

Critical weaknesses
  1. Limited methodological novelty and insufficient RAW encoding comparisons. The REC and RRE modules mainly combine established bit-plane processing, feature enhancement, and cross-modal alignment techniques. Although effective, the methodological novelty could be better clarified. In particular, the comparison with direct 8-bit compression may be insufficient to establish the necessity of bit-level encoding. Comparisons with normalized high-bit-depth RAW inputs and conventional RAW feature extractors would be helpful.

  2. Generalization across real low-light conditions. LOVS is collected using a single camera and a fixed 7-stop neutral-density filter. Although this provides controlled low-light conditions, it may not fully represent diverse real nighttime environments. Additional evaluations across cameras, exposure conditions, or naturally occurring low-light scenes would strengthen the generalization claims.

  3. Effectiveness of synthetic RAW supervision. The training RAW data are synthesized from normal-light RGB images through an inverse ISP pipeline. It remains unclear whether the improvements primarily come from RAW-specific information or the additional representation-learning supervision. A comparison with RGB-only feature distillation or alternative auxiliary supervision would better clarify the source of the gains.

  4. Further analysis of open-vocabulary generalization. Although OWL-Net achieves consistent improvements, some gains in mIoU are relatively modest. Given the substantial class imbalance in LOVS, a more detailed evaluation of seen versus unseen categories and frequent versus rare categories would help establish the effectiveness of the proposed method for open-vocabulary generalization.

Questions for authors

Please see the weaknesses part.

5 comments
AuthorsOpenReview ↗

Response Part 1

Dear Reviewer rgYX,

We sincerely thank you for recognizing the value of our dataset contribution and the effectiveness of our proposed framework. We also greatly appreciate your constructive comments and suggestions. Below, we provide detailed responses to address your concerns.

Q1: Novelty and the comparisons of RAW encoding strategies

Response: Thanks for your valuable comments. Regarding the methodological novelty, our contribution lies in the tailored design for RAW-assisted low-light open-vocabulary segmentation while maintaining RGB-only inference, which improves deployment flexibility and avoids the additional computational overhead. Specifically, REC explicitly models bit-level information by jointly considering bit variability and spatial coherence, allowing the model to emphasize informative structural details while suppressing noise-dominated bit planes. Meanwhile, RRE progressively enhances degraded RGB representations toward the RAW representation, enabling RGB-only inference without requiring additional RAW inputs.

Regarding RAW encoding comparisons, following your suggestion, we extend our experiments by introducing three additional baselines: (1) 10-bit RAW compression, which retains more intensity levels than the original 8-bit compression; (2) normalized 14-bit RAW, which directly processes high-bit-depth RAW values without explicit bit-plane decomposition; and (3) Dedicated RAW feature extractor, which replaces the REC module with the pretrained RAW-domain feature extractor from [R1]. For fair comparisons, we retain the same OpenWorldSAM framework and RRE module, replacing only the RAW encoding strategy. Meanwhile, all the configurations are re-trained on the same dataset (the synthetic low-light COCO-Stuff). The results on LOVS are summarized below.

RAW Encoding StrategymIoU ↑hIoU ↑
OpenWorldSAM (baseline)48.146.0
8-bit RAW compression47.648.9
10-bit RAW compression48.349.7
Normalized 14-bit RAW48.850.7
Dedicated RAW encoder [R1]48.350.4
REC (Ours)49.251.3

As shown in the table, increasing the quantization bit depth from 8 to 10 and 14 bits consistently improves the segmentation performance, highlighting the importance of preserving fine-grained RAW information. Directly using normalized 14-bit RAW or a dedicated RAW feature extractor also yields improvements over 8-bit compression. Nevertheless, our REC module achieves the best performance among all compared strategies, outperforming the normalized 14-bit RAW baseline by 0.4 mIoU and 0.6 hIoU. These results demonstrate that the performance gains do not simply come from retaining higher-bit-depth RAW values, but also benefit from our explicit bit-plane encoding and significance modeling.

Thanks for your suggestion. We have included the table with the corresponding analysis in our revised paper. (Presented as Table 4 in the main paper)

[R1] CVPR2025-Towards RAW Object Detection in Diverse Conditions

AuthorsOpenReview ↗

Response Part 2

Q2: Generalization across different low-light conditions

Response: We sincerely thank the reviewer for this valuable comment. We would like to clarify that our controlled acquisition protocol is primarily designed to ensure accurate pixel-level annotations and reliable evaluation for low-light open-vocabulary segmentation.

Specifically, annotating objects directly in extremely dark images is challenging, particularly for fine-grained categories and object boundaries. Therefore, we adopt a controlled acquisition strategy, where normal-light and low-light images are captured from the same viewpoint using an ND filter. This allows us to obtain accurate segmentation masks from the clearly visible normal-light images and directly transfer them to their spatially aligned low-light counterparts. Notably, similar ND-filter-based acquisition strategies have also been adopted in recent studies, including [R2] and [R3], to construct paired low-light and normal-light data.

Following the reviewer's suggestion, we further evaluate our method under a different attenuation level. During the construction of LOVS, we also retain images captured using a 6-stop ND filter, in addition to the original 7-stop setting. We conduct additional experiments on these images to examine the robustness of OWL-Net across different low-light conditions. The results are summarized below.

ND SettingMethodmIoU ↑hIoU ↑
6-stop (64×)OpenWorldSAM52.853.7
OWL-Net54.055.9
7-stop (128×)OpenWorldSAM48.146.0
OWL-Net49.251.3

As shown in the table, OWL-Net consistently outperforms OpenWorldSAM under both illumination conditions, achieving mIoU improvements of 1.2 and 1.1 under the 6-stop and 7-stop settings, respectively. These results further demonstrate the robustness of OWL-Net across different levels of low-light degradation.

Finally, we fully agree with the reviewer that broader real-world evaluations are important. In future work, we will further expand LOVS by incorporating images captured with different cameras, exposure conditions, and natural nighttime environments. We sincerely appreciate the reviewer's constructive suggestion, which provides valuable guidance for improving the diversity and practical applicability of our dataset.

Thanks for your valuable comment. We have added these results and the corresponding analysis to Appendix L (Table 12) of the revised manuscript.

[R2] CVPR24-Towards Robust Event-guided Low-Light Image Enhancement: A Large-Scale Real-World Event-Image Dataset and Novel Approach

[R3] ICCV25-Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring

Q3: Effectiveness of synthetic RAW supervision

Response: Thanks for your insightful comment. To address this concern, we conduct an additional experiment by replacing the synthetic RAW input to the REC module with the corresponding low-light RGB image. In this setting, the auxiliary branch extracts features from low-light RGB images rather than RAW data, while the RRE module continues to learn the enhanced RGB representations through the same feature alignment objective. For a fair comparison, we retain the same backbone, training configurations, and supervision strategy, with only the auxiliary input modality changed. The results on LOVS are summarized below.

MethodmIoU ↑hIoU ↑
OpenWorldSAM (baseline)48.146.0
OWL-Net (Low-light RGB to REC)48.750.6
OWL-Net (RAW to REC)49.251.3

The results demonstrate that RAW-guided supervision provides more effective guidance than RGB-based supervision, highlighting the benefit of the proposed RAW representation learning strategy beyond additional feature-level supervision alone. We appreciate the reviewer's constructive suggestion, which helps us provide a more comprehensive evaluation of the proposed framework. We have added the table with the corresponding analysis in our revised paper. (Presented as Table 3 in the main paper)

AuthorsOpenReview ↗

Response Part 3

Q4: Further analysis of open-vocabulary generalization

Response: Thank you for your valuable suggestion. To provide a more comprehensive evaluation of the open-vocabulary generalization capability of OWL-Net, we conduct additional experiments on LOVS from two perspectives: (1) seen vs. unseen categories and (2) frequent vs. rare categories.

(1) Seen vs. Unseen Categories. As described in Appendix I, among the 220 categories in LOVS, 14 categories overlap with the COCO-Stuff training label space and are considered seen categories, while the remaining 206 categories are regarded as unseen. To further investigate the generalization capability of OWL-Net, we separately evaluate the mIoU on seen and unseen categories. The experimental results are provided below.

MethodSeen mIoU ↑Unseen mIoU ↑
CAT-Seg26.028.2
OWL-Net (CAT-Seg)26.9 (+0.9)29.4 (+1.2)
OVRCOAT20.322.7
OWL-Net (OVRCOAT)25.5 (+5.2)27.7 (+5.0)
OpenWorldSAM43.948.3
OWL-Net (OpenWorldSAM)53.8(+9.9)48.9 (+0.6)

As shown in the table, OWL-Net consistently improves the segmentation performance on both seen and unseen categories across all three frameworks, demonstrating its effectiveness in enhancing low-light open-vocabulary segmentation. Specifically, OWL-Net improves the unseen mIoU of CAT-Seg and OVRCOAT by 1.2 and 5.0, respectively, which are comparable to the corresponding improvements on seen categories. For OpenWorldSAM, which already achieves relatively strong performance on unseen categories (48.3 mIoU), our method further improves the unseen mIoU by 0.6 points. These results demonstrate that OWL-Net can effectively enhance segmentation performance across different category groups and generalize to unseen semantic concepts under challenging low-light conditions.

(2) Frequent vs. Rare Categories. As illustrated in Fig. 13 of the Appendix, LOVS exhibits a long-tailed category distribution, where different categories have substantially different occurrence frequencies. To further examine the effectiveness of OWL-Net on infrequent categories, we divide the 220 categories into frequent and rare subsets based on their occurrence frequencies. Specifically, we rank all 220 categories according to their occurrence frequencies and designate the top 20% categories as frequent categories, while the remaining categories are considered rare categories. We then separately calculate the mIoU for each subset and compare OWL-Net with the corresponding baselines across the three frameworks.

MethodFrequent mIoU ↑Rare mIoU ↑
CAT-Seg26.028.5
OWL-Net (CAT-Seg)27.5 (+1.5)29.7 (+1.2)
OVRCOAT17.723.8
OWL-Net (OVRCOAT)21.7 (+4.0)29.0 (+5.2)
OpenWorldSAM39.150.3
OWL-Net (OpenWorldSAM)40.0 (+0.9)51.5 (+1.2)

As shown in the table, OWL-Net consistently improves the segmentation performance on both frequent and rare categories across all three frameworks, demonstrating its effectiveness under long-tailed category distributions. Specifically, OWL-Net achieves rare-category mIoU improvements of 1.2, 5.2, and 1.2 over CAT-Seg, OVRCOAT, and OpenWorldSAM, respectively. Notably, the improvements on rare categories are even greater than those on frequent categories for OVRCOAT and OpenWorldSAM. These results further demonstrate that OWL-Net can effectively enhance the segmentation of infrequently occurring categories, supporting its generalization capability under challenging low-light conditions.

These additional comparisons provide a more detailed assessment of the generalization capability of OWL-Net, particularly for unseen semantic categories and infrequently occurring objects. We appreciate the reviewer's constructive suggestion, which helps strengthen our experimental evaluation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Thanks, Contributions, New Experiments and New Clarifications

We sincerely thank all reviewers for their time and effort in evaluating our paper. We greatly appreciate their recognition of the practical significance of low-light open-vocabulary segmentation, the value of our LOVS dataset, and the consistent improvements achieved by OWL-Net across different OVS frameworks. We also appreciate the constructive suggestions, which motivate us to conduct extensive additional experiments and further strengthen our evaluation. Below, we summarize the key contributions recognized by the reviewers and our major updates during the rebuttal.

Contributions Recognized by Reviewers

  • Practical Significance. Our work addresses an important yet underexplored problem of low-light open-vocabulary segmentation. The RAW-assisted training strategy with RGB-only inference provides practical value and maintains low deployment costs. [rgYX, 9Shc, ZwzA]
  • Valuable Dataset Contribution. Our LOVS dataset provides 1,000 aligned normal-/low-light image pairs, covering 220 semantic categories and 6,949 object-level masks. The reviewers recognize its potential as a valuable benchmark for the research community. [rgYX, 9Shc, ZwzA]
  • Effectiveness and Generality. OWL-Net consistently improves three structurally different OVS frameworks (CAT-Seg, OVRCOAT, and OpenWorldSAM) across synthetic and real-world low-light benchmarks, demonstrating its plug-and-play capability and practical effectiveness. [rgYX, 9Shc, ZwzA]

Additional Experiments and Clarifications

Following the reviewers' suggestions, we conduct extensive additional experiments and analyses to directly address the major concerns, with the key updates summarized below.

New Experiments:

  • Effectiveness of RAW Supervision. We introduce additional ablations by removing RAW guidance, replacing RAW inputs with normal-/low-light RGB images, and removing the RAW-to-RGB alignment loss. The results consistently demonstrate the benefits of RAW-assisted feature learning beyond additional network components or supervision alone. [rgYX, 9Shc, ZwzA]
  • Comparisons of RAW Encoding Strategies. We introduce additional baselines, including 10-bit RAW compression, normalized 14-bit floating-point RAW, and a dedicated pretrained RAW encoder. Our REC module consistently achieves the best performance, further supporting the effectiveness of the proposed bit-level encoding strategy. [rgYX, ZwzA]
  • Generalization Across Low-Light Conditions. We extend the evaluation to images captured with a 6-stop ND filter (64× attenuation), in addition to the original 7-stop setting (128×). OWL-Net consistently outperforms OpenWorldSAM under both conditions. We also investigate the effects of different synthetic training attenuation levels. [rgYX, 9Shc, ZwzA]
  • Evaluation Reliability and Open-Vocabulary Generalization. We conduct repeated experiments with three random seeds, demonstrating stable improvements with low performance variance. We further analyze segmentation performance on seen/unseen and frequent/rare categories to provide a more comprehensive assessment of open-vocabulary generalization. [rgYX, ZwzA]
  • Accuracy–Efficiency Trade-off. We introduce the lightweight enhancement method SCI for additional two-stage comparisons. OWL-Net achieves better segmentation accuracy and lower inference latency than SCI + OpenWorldSAM, further demonstrating its favorable balance between accuracy and efficiency. [ZwzA]
  • Low-Light Synthesis Fidelity. We introduce three additional evaluation metrics (SSIM, FID, and LPIPS) alongside PSNR on both LOVS and LIS, providing a more comprehensive quantitative evaluation of our low-light synthesis pipeline. [ZwzA]

Further Clarifications:

  • Training Fairness and Reproducibility. We provide detailed descriptions of the pretrained checkpoints, trainable modules, and training configurations for each baseline and its OWL-Net counterpart, clarifying the fairness of our comparisons. [9Shc, ZwzA]
  • Dataset Construction and Related Work. We further clarify the rationale behind our ND-filter-based acquisition strategy, particularly its importance for reliable pixel-level annotations. Meanwhile, we discuss the distinctions between our approach and existing low-light perception and RAW-guided methods. [rgYX, 9Shc, ZwzA]

We sincerely appreciate the reviewers' valuable feedback. We believe these additional experiments and clarifications substantially strengthen the experimental evidence, evaluation reliability, and practical contributions of our work. We have incorporated many of these results into the revised manuscript and provided detailed point-by-point responses. We hope our efforts adequately address the reviewers' concerns and support a positive reassessment of our submission.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.