acceptodds
Under review as a conference paper at ICLR 2027

Intrinsic Feature Decomposition for Domain Adaptive Low-Light Object Detection

Abstract

Object detection under low-light conditions remains underexplored due to the challenge of learning discriminative representations across extreme illumination variations. We propose a domain-adaptive pretraining framework that learns noise-robust, illumination-invariant feature representations directly within the detector via feature-level decomposition. Inspired by Retinex theory, our approach treats the first-layer backbone feature as intrinsic structure and employs an auxiliary encoder-decoder during pretraining to factor features into intrinsic, illumination, and noise components. The auxiliary branch is discarded after pretraining, incurring no inference overhead. Using paired normal-dark images synthesized from COCO, we jointly optimize detection and decomposition losses to enforce intrinsic consistency. When fine-tuned on real-world low-light datasets, our pretraining consistently improves the COCO-pretrained baseline, with gains of 7.9 mAP on LOD, 4.2 on NOD, 3.7 on NuImages, 2.7 on BDD100K, 11.7 on ExDark, and 5.9 on DARK Face. These gains persist across detector architectures, including YOLO26-m and RT-DETR. Overall, these results show that embedding Retinex principles at the feature level effectively bridges the illumination gap for low-light detection, scales across model sizes and detector architectures, and preserves runtime efficiency. Our code will be publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Object Detection, Domain Adaptation, Robust Perception

Reviews

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

The authors propose an intrinsic feature decomposition framework for low-light object detection. The approach decomposes detector features into intrinsic, illumination, and noise components during pretraining, encouraging illumination-invariant representation learning. The authors also introduce auxiliary decomposition modules that are removed after pretraining. Therefore, they introduce no additional inference overhead. Experiments across multiple benchmarks and detector architectures demonstrate consistent improvements over standard pretraining models and existing baselines.

Critical strengths
  1. The proposed method achieves consistent improvements on multiple low-light object detection benchmarks and also performs well on nighttime driving benchmarks. Its effectiveness across YOLO variants and RT-DETR further supports the approach's generalizability.

  2. The feature decomposition is introduced only during pretraining, allowing the detector to benefit from illumination-robust representations without additional inference-time computation. This makes the framework practical for real-world applications where inference efficiency and speed are important.

Critical weaknesses
  1. The proposed approach builds upon Retinex decomposition and illumination-invariant representation learning. Previous studies such as MAET [1] and DAI-Net [2] have applied similar ideas for low-light object detection. Although applying decomposition directly to the detector's feature space is a meaningful design choice, the paper could more clearly distinguish the conceptual innovation beyond these closely related approaches.

  2. The ablation study provides valuable information for the effectiveness of feature-level decomposition. However, it is still unclear how much of the improvement comes from the proposed Retinex-inspired formulation, instead of the additional introduced reconstruction supervision during pretraining. A comparison with a similarly sized auxiliary reconstruction objective without explicit decomposition would help answer this question and strengthen the innovation.

[1]. https://openaccess.thecvf.com/content/ICCV2021/papers/Cui_Multitask_AET_With_Orthogonal_Tangent_Regularity_for_Dark_Object_Detection_ICCV_2021_paper.pdf [2]. https://openaccess.thecvf.com/content/CVPR2024/papers/Du_Boosting_Object_Detection_with_Zero-Shot_Day-Night_Domain_Adaptation_CVPR_2024_paper.pdf

Questions for authors
  1. How much additional computational cost (e.g., GPU hours or FLOPs) is introduced by the auxiliary decomposition modules during pretraining, compared with the vanilla model?
4 comments
AuthorsOpenReview ↗

Official Comment to Reviewer hyEs (1/3)

Dear Reviewer hyEs,

We sincerely thank the reviewer for the careful review of our paper and for the positive assessment of the consistent improvements across multiple benchmarks and detector architectures, as well as the practicality of a design that adds no inference cost. In particular, the suggestion to isolate the source of the gains was of great help in verifying the central claim of our paper. We respond to each point below.

Q1: Conceptual Innovation That Distinguishes Our Work from MAET and DAI-Net

We appreciate the reviewer's assessment that applying the decomposition directly in the detector's feature space is a meaningful design choice. To explain the differences from prior work more clearly, we have added a conceptual comparison of the training pipelines of MAET, DAI-Net, and our method (Fig. 8) to Appendix D of the revised version, and we summarize the differences along three axes.

  1. Where the decomposition is defined.
    • MAET employs a self-supervised auxiliary task that predicts the degradation transformation and separates object information from degradation information through an orthogonality constraint. The target of the auxiliary task is the degradation transformation, and neither a Retinex structure nor image reconstruction is used.
    • DAI-Net attaches a decoder to the shallow features of the detector to recover an image-level reflectance. The products of the decomposition (reflectance and illumination) are defined in the image space, and the detector features serve as the input to that decoder. DAI-Net also includes a procedure that interchanges reflectance, but this too is carried out in the image space.
    • We treat the first-layer feature of the detector backbone itself as the factor I of the decomposition and jointly learn an illumination factor L and a noise feature N of the same dimensions, forming a factorization in the feature space, F = (I + N) ⊙ L. What is exchanged in cross-reconstruction is also the detector feature itself.
  2. Source of the supervision signal. The reflectance target of DAI-Net is provided by a separately pretrained decomposition network, so the quality of that network bounds the invariance that can be learned. In our own examination, we observed that the pretrained decomposition network collapses on some target datasets (Fig. 9 in Appendix D.1 of the revised version). In contrast, we learn the decomposition solely through reconstruction and cross-reconstruction of the input images, without any external decomposition network or pseudo ground truth.
  3. Explicit modeling of noise. Under low light, noise degradation that cannot be explained by illumination modulation alone is substantial. MAET and DAI-Net have no corresponding term, whereas we introduce N explicitly so that the noise is not absorbed into I. This term adds +1.2 mAP on LOD (from 62.5 to 63.7).

These conceptual differences translate into differences in performance. With the same detector, the same Dark ISP synthesis, and the same fine-tuning schedule, the proposed method consistently outperforms MAET and DAI-Net.

MethodLODNODNuImages (night)BDD100K (night)ExDarkDARK Face
MAET61.150.430.725.673.568.9
DAI-Net61.450.531.325.674.768.8
Ours63.752.532.727.079.171.5

Moreover, within our framework, the variant that defines the decomposition in the image space reaches only 57.2 mAP, whereas defining it in the feature space reaches 62.5 mAP (Table 3). This result shows that "where the decomposition is defined" is the key.

AuthorsOpenReview ↗

Official Comment to Reviewer hyEs (2/3)

Q2: Whether the Gains Come from the Retinex-Based Formulation or from the Added Reconstruction Supervision

We thank the reviewer for the careful and constructive feedback. As suggested, we conducted a control experiment that provides only an auxiliary reconstruction task of matched capacity, without the decomposition structure. In all settings, pretraining is conducted with the same Dark ISP pairs and the same schedule, and the resulting detectors are subsequently fine-tuned on LOD.

This control uses the same auxiliary encoder and decoder as our method, but simply adds the encoder output to the first-layer feature before reconstructing the input image (D(I + E(S))). It has no multiplicative structure, no distinction between L and N, and no smoothness constraint, and the number of parameters of the auxiliary modules is identical to that of our method. In other words, it provides reconstruction supervision with the same capacity but without the decomposition structure.

Table R11. Separating the contributions of reconstruction supervision and the decomposition structure (LOD, YOLOv8-m)

Pretraining settingAuxiliary moduleDecomposition structuremAP
COCO baseline––55.8
Dark ISP pair training only––58.1
↳ + paired feature alignment (consistency) only––58.4
↳ + capacity-matched reconstruction (no decomposition structure)Encoder + decoderNone59.0
↳ + image-level Retinex decompositionEncoder + decoderImage space57.2
↳ + feature-level decomposition (without N)Encoder + decoderFeature space62.5
↳ + feature-level decomposition + N (Ours)Encoder + decoderFeature space63.7

The results can be summarized as follows.

  • Reconstruction supervision by itself is helpful, but its gain is limited to +0.9 over Dark ISP pair training (from 58.1 to 59.0).
  • Compared with this control, whose auxiliary-module capacity is matched exactly to ours (59.0), the feature-level decomposition structure brings an additional +4.7.

Therefore, most of the improvement comes not from the added reconstruction supervision but from the Retinex-style decomposition structure in which the detector feature serves as a factor.

We also examined the role of each loss term within the decomposition structure through additional ablation.

Table R12. Loss-term ablation (LOD, YOLOv8-m)

SettingmAP
Ours (full)63.7
Without the noise term N62.5
Without the cross-reconstruction loss63.0
Without the illumination smoothness loss61.9
Without the intrinsic consistency loss59.3

Every term contributes to the performance. Removing the intrinsic consistency loss causes the largest drop (−4.4 mAP), followed by the illumination smoothness loss (−1.8), the noise term (−1.2), and the cross-reconstruction loss (−0.7). In particular, removing the smoothness loss allows the illumination factor L to absorb structural information, which reduces the role of I, and removing the consistency loss eliminates the direct constraint that aligns I across the pair. The fact that performance drops in both cases supports that the gains arise when the decomposition behaves as intended. We are grateful to the reviewer for the constructive comment, which allowed us to further strengthen our work through this control experiment and the additional ablations.

AuthorsOpenReview ↗

Official Comment to Reviewer hyEs (3/3)

Q3: Additional Pretraining Cost Introduced by the Auxiliary Decomposition Modules

As the reviewer notes, the auxiliary encoder-decoder increases the computation of the pretraining stage, which we also stated as a limitation in the Conclusion of the submission. To quantify this cost, we measured the parameters and computation of the three auxiliary architectures compared in Appendix A.4 (Table 6).

Table R8. Additional pretraining cost by auxiliary encoder-decoder architecture (YOLOv8-m, per 640×640 input image; LOD mAP from Table 6 in the appendix of the submission)

Auxiliary encoder-decoderParameters (enc + dec)Relative to the detector (25.9M)GFLOPs (enc + dec)Relative to the detector (78.9)LOD mAP
MLP-based41.6K (32.0K + 9.6K)0.16%8.38 (6.37 + 2.01)10.6%63.3
Convolutional (Ours)89.5K (67.4K + 22.1K)0.35%19.05 (13.74 + 5.31)24.1%63.7
Light Transformer-based146.6K (105.5K + 41.1K)0.57%37.57 (26.09 + 11.48)47.6%64.3

The results can be summarized as follows.

  • Magnitude of the additional cost. The three-layer convolutional architecture that we use has 89.5K parameters (0.35% of the detector) and requires 19.05 GFLOPs per image (24.1% of the detector). The computation is large relative to the parameter count because the auxiliary modules operate at the same high resolution as the first-layer feature of the backbone.
  • Cost and performance across architectures. The computation varies from 10.6% to 47.6% of the detector depending on the architecture, whereas the LOD mAP differs only slightly, ranging from 63.3 to 64.3. The convolutional architecture that we adopt is therefore a reasonable choice that offers a favorable balance between performance and cost, as discussed in Appendix A.4. When the pretraining cost is more critical, the MLP architecture can be used instead, which reduces the computation to less than half at a cost of only 0.4 mAP.
  • The cost is incurred only once, during COCO pretraining. Starting from COCO-pretrained weights and fine-tuning them on multiple target datasets is a widely adopted practice for training detectors, and our pretrained weights can be used in exactly the same manner. We performed pretraining only once per detector and used the same weights for all six target datasets. We also plan to release the code and the pretrained weights, so that users of the released weights will not need to bear this cost again.
  • There is no additional cost in fine-tuning or inference. The auxiliary encoder-decoder is removed after pretraining. Fine-tuning therefore proceeds at the same cost as with the standard COCO-pretrained weights, and the parameters (25.9M), computation (78.9 GFLOPs), and latency at inference are identical to those of the baseline.

In summary, the additional cost is confined to a one-time pretraining, and the subsequent fine-tuning and inference are performed at the same cost as the baseline.

We thank the reviewer once again for the constructive comments. We hope that the additional experiments and clarifications have helped to address the concerns, and we will incorporate all of them into the final version of the paper.

Dear Reviewer,

The authors have responded to your review. Could you please take a moment to review their responses and engage in the discussion?

Thank you for your time and valuable contributions.

Kind regards, AC

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

This paper proposes a domain-adaptive pretraining framework based on Retinex theory to improve object detection performance in extremely low-light conditions.

The proposed method performs Intrinsic Feature Decomposition at the feature level rather than the image pixel level, utilizing the first-layer backbone feature of the object detector as the intrinsic structure.

During the pretraining phase, an auxiliary encoder-decoder is used to decompose the input features into intrinsic, illumination, and noise components.

It jointly optimizes detection and decomposition losses using paired normal-light and synthetic dark images generated from the COCO dataset.

After pretraining, the auxiliary modules are completely discarded during finetuning and actual inference, introducing no additional computational overhead.

Critical strengths

Broad Performance Improvement & Generalization: The proposed method is applicable to various model architectures (YOLOv8-m, YOLO26-m, RT-DETR) and consistently outperformed existing baselines across multiple benchmarks and real-world night driving datasets, including LOD (+7.9 mAP), NOD (+4.2 mAP), NuImages (+3.7 mAP), BDD100K (+2.7 mAP), ExDark, and DARK Face.

Zero Inference Overhead: Since the auxiliary encoder and decoder networks for feature decomposition are used solely during the pretraining phase and discarded afterward, they do not introduce any inference latency, which is crucial for real-time object detection tasks.

Scalability with Model Capacity: While standard COCO-pretrained baselines suffer from performance saturation in low-light conditions as model capacity increases (from YOLOv8-s onwards), the proposed pretraining strategy demonstrated excellent scalability, showing proportional and steady performance gains as the model parameter size increases (from YOLOv8-n to YOLOv8-x).

Critical weaknesses

Synthetic-to-Real Gap: Since the model's pretraining relies entirely on synthetic dark images generated via the Dark ISP pipeline, the gap between synthetic data and real-world environments remains a limitation.

Additional Computational Cost During Pretraining: Although there is no overhead during inference, the pretraining process requires running and optimizing the auxiliary encoder-decoder modules alongside the main network, which adds computational cost during training.

Questions for authors

If the pretraining were conducted directly using paired real-world normal/low-light datasets (e.g., LOD paired data) instead of synthetic image pairs from the Dark ISP pipeline, would it achieve a similar level of domain adaptation compared to the synthetic data-based learning?

Instead of completely discarding the auxiliary modules during inference, have you explored the accuracy-latency tradeoff of retaining an extremely lightweight illumination estimation module to dynamically adjust features in response to unpredictable real-time illumination changes?

6 comments
AuthorsOpenReview ↗

Official Comment to Reviewer farQ (1/5)

Dear Reviewer farQ,

We sincerely thank the reviewer for the positive assessment of our paper. We appreciate the recognition of the consistent improvements across multiple detectors and benchmarks, the design that incurs no inference cost, and the scalability with model capacity. The two questions raised by the reviewer also gave us a valuable opportunity to examine the scope of applicability of our method more broadly. We respond to each point below.

Q1: Gap between Synthetic Data and Real-World Environments

We thank the reviewer for raising this important point. The gap pointed out by the reviewer does exist. It is not specific to our method; it is a limitation shared by the entire line of work that learns illumination robustness from low-light images synthesized with Dark ISP, such as MAET and DAI-Net, and we also acknowledged it as a limitation in the Conclusion of the submission. Real-world degradations that the synthesis pipeline does not model are not directly addressed during pretraining.

Given this limitation, what we aimed to show is that, even when starting from the same synthetic data, the proposed pretraining generalizes to real data most broadly within this line of work. For this reason, all evaluations were conducted on real captured low-light images (synthetic images are used only for pretraining), and we extended the scope to six benchmarks of different characteristics and three detector architectures. During the rebuttal period, we examined this point more concretely in the following three ways.

(1) Comparison with methods that use the same Dark ISP synthetic pairs. Since the synthetic data are identical, the differences arise only from what is learned from those data. The submission provided this comparison only on LOD and NOD, and we have now added MAET and DAI-Net on the remaining four benchmarks as well.

Table. Performance on real benchmarks of methods pretrained with Dark ISP synthetic pairs (YOLOv8-m, after fine-tuning, mAP; excerpted from Table 1 and Table R3)

MethodLODNODNuImages (night)BDD100K (night)ExDarkDARK Face
Baseline (no synthetic pairs)55.848.329.024.367.465.6
MAET61.150.430.725.673.568.9
DAI-Net61.450.531.325.674.768.8
Ours63.752.532.727.079.171.5

The proposed method achieves the highest performance on all six benchmarks. In other words, even with identical synthetic data, generalization to real data depends on what is learned from them, and the proposed feature-level decomposition transfers what is learned from synthetic data to real environments more effectively than the compared methods.

AuthorsOpenReview ↗

Official Comment to Reviewer farQ (2/5)

(2) Analysis by illumination type. ExDark provides ten real illumination types. Dark ISP models only a global reduction in illumination and sensor noise, and does not directly model non-uniform illumination such as strong light sources, screens, or windows. We examined whether the gains are maintained for such types as well.

Table R6. Performance by illumination type on ExDark (YOLOv8-m, after fine-tuning, mAP50)

Illumination typeAllLowAmbientObjectSingleWeakStrongScreenWindowShadowTwilight
Baseline67.478.667.166.672.462.264.767.173.184.185.7
Ours79.188.384.473.283.676.575.086.780.291.389.7
Gain+11.7+9.7+17.3+6.6+11.2+14.3+10.3+19.6+7.1+7.2+4.0

The performance improves for all ten illumination types. For Strong, Screen, and Window, which are non-uniform illumination types that Dark ISP does not directly model, the improvements are +10.3, +19.6, and +7.1, respectively, which indicates that the gains are not limited to the global low-light condition covered by the synthesis pipeline. In other words, what the proposed pretraining learns is not a specific degradation pattern produced by Dark ISP but a representation that is robust to the diverse illumination variations of the real world, and these results explicitly confirm that the gains are maintained under real illumination conditions that are not covered by the synthesis process.

(3) Consistency of the intrinsic feature on real pairs. The analysis in Section 4.5 compares the consistency of the intrinsic feature of our method under synthetic degradation. To address the reviewer's concern about the synthetic-to-real gap, we additionally measured the consistency on the real long-/short-exposure pairs of LOD.

Table R7. Feature consistency on real pairs (LOD long-/short-exposure)

MethodPearson correlationCosine similarity
Baseline0.5610.634
MAET0.6040.649
DAI-Net0.6890.721
Ours0.8340.845

On real pairs, the feature consistency of the proposed method is 0.834 in Pearson correlation and 0.845 in cosine similarity, which is higher than that of MAET (0.604, 0.649) and DAI-Net (0.689, 0.721), both of which are trained with the same synthetic data. In other words, the illumination invariance learned from synthetic pairs is not fitted only to synthetic degradation; it also holds for the illumination differences that arise in real captures.

In summary, the synthetic-to-real gap remains a limitation common to the line of work based on Dark ISP, and our method is no exception. Nevertheless, among the methods that use the same synthetic data, the proposed method generalizes to real data most consistently, and its effect is maintained for illumination types that Dark ISP does not directly model and for real captured pairs. We are grateful to the reviewer for this comment, which allowed us to address the concern about the synthetic-to-real gap and to make our claims more solid.

AuthorsOpenReview ↗

Official Comment to Reviewer farQ (3/5)

Q2: Additional Computational Cost during Pretraining

As the reviewer notes, the auxiliary encoder-decoder increases the computation of the pretraining stage, which we also stated as a limitation in the Conclusion of the submission. To quantify this cost, we measured the parameters and computation of the three auxiliary architectures compared in Appendix A.4 (Table 6).

Table R8. Additional pretraining cost by auxiliary encoder-decoder architecture (YOLOv8-m, per 640×640 input image; LOD mAP from Table 6 in the appendix of the submission)

Auxiliary encoder-decoderParameters (enc + dec)Relative to the detector (25.9M)GFLOPs (enc + dec)Relative to the detector (78.9)LOD mAP
MLP-based41.6K (32.0K + 9.6K)0.16%8.38 (6.37 + 2.01)10.6%63.3
Convolutional (Ours)89.5K (67.4K + 22.1K)0.35%19.05 (13.74 + 5.31)24.1%63.7
Light Transformer-based146.6K (105.5K + 41.1K)0.57%37.57 (26.09 + 11.48)47.6%64.3

The results can be summarized as follows.

  • Magnitude of the additional cost. The three-layer convolutional architecture that we use has 89.5K parameters (0.35% of the detector) and requires 19.05 GFLOPs per image (24.1% of the detector). The computation is large relative to the parameter count because the auxiliary modules operate at the same high resolution as the first-layer feature of the backbone.
  • Cost and performance across architectures. The computation varies from 10.6% to 47.6% of the detector depending on the architecture, whereas the LOD mAP differs only slightly, ranging from 63.3 to 64.3. The convolutional architecture that we adopt is therefore a reasonable choice that offers a favorable balance between performance and cost, as discussed in Appendix A.4. When the pretraining cost is more critical, the MLP architecture can be used instead, which reduces the computation to less than half at a cost of only 0.4 mAP.
  • The cost is incurred only once, during COCO pretraining. Starting from COCO-pretrained weights and fine-tuning them on multiple target datasets is a widely adopted practice for training detectors, and our pretrained weights can be used in exactly the same manner. We performed pretraining only once per detector and used the same weights for all six target datasets. We also plan to release the code and the pretrained weights, so that users of the released weights will not need to bear this cost again.
  • There is no additional cost in fine-tuning or inference. The auxiliary encoder-decoder is removed after pretraining. Fine-tuning therefore proceeds at the same cost as with the standard COCO-pretrained weights, and the parameters (25.9M), computation (78.9 GFLOPs), and latency at inference are identical to those of the baseline.

In summary, the additional cost is confined to a one-time pretraining, and the subsequent fine-tuning and inference are performed at the same cost as the baseline.

Reviewer_aUY2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose a new framework for low-light object detection. Based on Retinex theory, the authors use the output of the first convolutional layer of the decoder as I, an auxiliary encoder to predict L and N, and an auxiliary shared decoder to map these features back to images. To train this framework, the authors use the Dark ISP pipeline in MAET to synthesize low-light pairs. The training objectives are listed in Section 3.4.

Critical strengths
  1. The authors provide experiments on two datasets (LOD, NOD). They also provide extra experiments against the naive baseline on an extra night driving task and two more low-light detection datasets (ExDark, DARK Face).
  2. The authors provide an analysis (Section 4.5) to demonstrate that the proposed feature is more robust to illumination compared with the baseline.
Critical weaknesses
  1. The framework might have limited novelty. Introducing Retinex theory and feature decomposition is not new for low-light enhancement. The minor variations in Eqs. (3)–(4) and the simple framework design do not provide sufficient novelty for an ICLR paper. The loss function designs are also common in low-light enhancement. Moreover, the training data are synthesized using an existing strategy.

  2. In Table 1, all compared methods are from before 2026. The authors should add more recent methods for comparison.

  3. For the compared domain adaptation methods, many of them are zero-shot, which means they do not require low-light detection annotations. If the proposed method requires low-light detection annotations, outperforming these zero-shot domain adaptation methods is not a fair comparison.

  4. I understand that the results on night driving scenes, ExDark, and DARK Face are extra. However, simply outperforming the baseline is not sufficient to demonstrate the superiority of the proposed method.

  5. From Figure 5, we can see that the proposed features are still relevant to degradation and are not fully stable.


[Update] The authors' response does not resolve my concerns, particularly those regarding the technical contribution. Low-light image enhancement and low light object detection do not differ much. Decomposing Retinex features is not sufficient for an ICLR paper.

Questions for authors
  1. Please provide more details about "All models are fine-tuned on the LOD/NOD." As far as I am concerned, for many compared methods, their training strategies and hyperparameters are designed for their original datasets. How do the authors address this gap?

  2. "After pretraining, the auxiliary encoder and decoder are discarded, and the pretrained detector is fine-tuned on real low-light datasets without additional inference cost." Does the proposed method require low-light object detection images and annotations during fine-tuning?

5 comments
AuthorsOpenReview ↗

Official Comment to Reviewer aUY2 (1/4)

Dear Reviewer aUY2,

We sincerely thank the reviewer for the careful review of our paper and for the positive assessment of the experiments on multiple benchmarks and the feature robustness analysis in Section 4.5. The points raised by the reviewer have allowed us to make the comparison protocol and the scope of our contribution considerably clearer. We respond to each point below.

Q1: Novelty and Differences from Existing Retinex-Based Methods

We would first like to clarify the problem that our work addresses. Our method is not a low-light image enhancement (LLIE) method; it is a detector pretraining method for low-light object detection. Retinex-based enhancement methods decompose an image and output a visually improved image. In our method, by contrast, what is decomposed is not the image but the features of the detector, and the outcome is not an enhanced image but a pretrained detector. Our contribution therefore does not lie in adding a noise term to the Retinex equation. Rather, it lies in reformulating a Retinex-inspired decomposition at the feature level within detector pretraining, so that the detector learns illumination-invariant intrinsic features that directly benefit downstream detection.

Retinex theory, loss forms such as reconstruction and smoothness, and Dark ISP synthesis have, in part, been used in prior work, and we do not claim to have newly proposed these components. However, when these components are applied at the image level as in existing approaches, improvements in visual quality do not translate into detection performance. This is directly confirmed by the experimental results in our paper.

  • Detection after enhancement (Table 1). Even when seven LLIE methods are applied, the mAP on LOD ranges from 55.3 to 56.8, which remains within −0.5 to +1.0 of the baseline without enhancement (55.8). Among them, the Retinex-based methods RUAS, SCI, and Retinexformer achieve 56.6, 55.5, and 55.6, respectively.
  • Image-level Retinex decomposition as an auxiliary task (Table 3). When the Retinex decomposition is placed in the image space within our own framework, the mAP is 57.2, which is even lower than that obtained by pretraining with Dark ISP pairs alone (58.1).

The key points of our design are the following four.

  1. The feature that the detector uses at inference is itself a factor of the decomposition. Existing Retinex-based enhancement methods decompose the image, and DAI-Net decodes an image-level reflectance from the detector features. We instead treat the first-layer feature of the backbone as the intrinsic factor I and learn a factorization in the feature space, F = (I + N) ⊙ L. The invariance constraint is therefore imposed directly on the representation that is actually used at inference. (For clarity, I denotes the output of the first convolutional layer of the detector backbone, not that of the decoder.)
  2. No external supervision is required. The decomposition is learned solely through reconstruction and cross-reconstruction, without a pretrained decomposition network, reflectance pseudo ground truth, or ground-truth enhanced images.
  3. Noise is modeled explicitly. The noise feature N accounts for the low-light degradation that cannot be explained by illumination modulation alone, which prevents this degradation from being absorbed into I.
  4. The inference architecture is unchanged. The auxiliary modules consist only of an encoder and a decoder, each with three convolutional layers, and they are removed after pretraining. Approaches that first enhance the input must run an enhancement network at every inference, whereas the inference cost of our method is identical to that of the baseline.

To make these differences easier to compare, we have added a conceptual comparison of the training pipelines of Retinex-based LLIE, MAET, DAI-Net, and our method to Appendix D of the revised version (Fig. 8). In addition, Appendix D.1 (Fig. 9) shows that the pretrained decomposition network on which DAI-Net relies works well on LOL, its training data, but fails on NuImages, which supports the advantage of a design that does not depend on an external decomposition network.

Regarding the synthetic data, we deliberately used the same Dark ISP as in the protocols of MAET and DAI-Net. Keeping the synthesis identical allows the effects of the training objective and the pretraining framework to be isolated and compared fairly. The proposed method is not tied to any particular low-light synthesis scheme. Should the reviewer have any further questions regarding our contribution, we would be glad to address them during the discussion phase.

AuthorsOpenReview ↗

Official Comment to Reviewer aUY2 (2/4)

Q2: Comparison Only with Methods Published before 2026

As suggested, we have additionally compared our method with two recent methods. FRBNet (Sun et al., NeurIPS 2025) is a plug-and-play module that extends the Lambertian assumption of YOLA to the frequency domain. AMIEOD (Huang et al., TMM 2026) jointly trains multiple image enhancement modules with the detector and, at inference, uses a selection module to choose and apply one of them for each input. Both methods were implemented on YOLOv8-m and trained with the same fine-tuning schedule as in Table 1.

Table R1. Comparison with recent methods (YOLOv8-m, mAP)

MethodVenueLODNODExDarkDARK Face
Baseline–55.848.367.465.6
YOLANeurIPS 202461.250.972.069.2
FRBNetNeurIPS 202557.348.972.869.6
AMIEODTMM 202657.650.375.369.4
Ours–63.752.579.171.5

The proposed method outperforms both recent methods on all four benchmarks. We are grateful to the reviewer for this constructive comment, as the additional baseline experiments have further strengthened our paper.

Q3: Fairness of the Comparison with Zero-Shot Methods, and Whether Fine-Tuning Requires Target Labels (Weakness 3, Question 2)

We first clarify the original training settings of the compared methods and then explain our comparison protocol.

(1) Most of the compared domain adaptation methods are not zero-shot methods. LLIE methods are preprocessing steps that enhance the input image, and MAET and YOLA are also trained with the labels of the target dataset in their original papers. DAI-Net is the only method that considers a zero-shot setting, and, as shown in (2) below, the original DAI-Net paper does not evaluate the method only in the zero-shot setting.

(2) Our comparison with DAI-Net follows the protocol of the original DAI-Net paper. The experiments of DAI-Net consist of two settings.

  • Zero-shot setting. This setting is used when the source and the target share the same classes.
  • COCO pretraining followed by target fine-tuning. When DAI-Net is compared with MAET using a general object detector (YOLOv3), the detector is pretrained on COCO and then fine-tuned with the training split and labels of the target dataset.

Our comparison corresponds to the second setting. This "COCO pretraining → target fine-tuning" protocol was introduced by MAET and followed by DAI-Net, and it is widely used in domain adaptation research for low-light detection. Therefore, we did not compare DAI-Net under a condition that the method was not designed for; we compared it under the very condition that the original DAI-Net paper uses for its comparison with MAET.

Table R2. Training settings in the original papers and comparison settings in our paper

MethodTraining setting in the original paperComparison setting in our paper
LLIE methodsImage enhancement network (separate from detector training)COCO-pretrained detector; inputs enhanced with the official weights; fine-tuned with target labels
MAETCOCO pretraining → fine-tuning with target labelsSame as the original paper
DAI-Net(a) Zero-shot: trained on a source with the same classes and evaluated on the target; (b) COCO pretraining → fine-tuning with target labelsSame as (b)
YOLASupervised training with target labelsSame as the original paper
Ours–COCO pretraining (with the decomposition loss) → fine-tuning with target labels

(3) Therefore, all compared methods in Table 1, including ours, use the training images and box labels of the target low-light dataset in exactly the same way during fine-tuning. All methods are implemented on the same detector (YOLOv8-m) and are fine-tuned with the same target training split and labels under the same schedule. The only difference among the methods is the pretraining strategy (or module) on COCO. In other words, no method in Table 1 is evaluated in a zero-shot manner, and Table 1 is a fair comparison in which every method uses the target labels under the same supervision.

If anything remains unclear or the reviewer has further questions, we would be glad to clarify.

AuthorsOpenReview ↗

Official Comment to Reviewer aUY2 (3/4)

Q4: Comparison Only with the Baseline on Night Driving, ExDark, and DARK Face

We agree with the reviewer. Outperforming the baseline alone is not sufficient to demonstrate the superiority of a method. In the submission, these experiments were intended to verify generalization. Following the reviewer's comment, we have added MAET, DAI-Net, and YOLA to all four benchmarks. All methods use YOLOv8-m and follow the same fine-tuning schedule for each dataset.

Table R3. Comparison with domain adaptation methods on the additional benchmarks (YOLOv8-m, mAP)

MethodNuImages (night)BDD100K (night)ExDarkDARK Face
Baseline29.024.367.465.6
MAET30.725.673.568.9
DAI-Net31.325.674.768.8
YOLA30.925.272.069.2
Ours32.727.079.171.5

The proposed method achieves the highest performance on all four benchmarks, exceeding the second-best method on each benchmark by +1.4, +1.4, +4.4, and +2.3, respectively. The comparison with the recent methods (FRBNet and AMIEOD) is provided in Table R1 of Q2.

Q5: The Features in Fig. 5 Are Still Affected by Degradation

In the paper, we describe this behavior as a "gradual decrease" (L412) and do not claim complete invariance. What we intended to show through Figures 5 and 6 is that our features exhibit relatively small degradation compared with the baseline.

To address the reviewer’s further concern, we added MAET and DAI-Net to the analysis of Fig. 5 under the same protocol and compared the magnitude of the degradation (Table R4). We have replaced Figure 5 and Table 8 in the revised version with these updated results, and we kindly ask the reviewer to refer to them.

Table R4. Robustness of the first-layer (P1) feature at each degradation level (COCO val; Pearson correlation ρ and cosine similarity s with respect to the original image, mean ± std; α and β of each level are given in Table 5 of the appendix of the submission; the values of Baseline and Ours are from Table 8 of the submission, and MAET and DAI-Net are measured under the same protocol)

Degradation level (α, β)Baseline ρMAET ρDAI-Net ρOurs ρBaseline sMAET sDAI-Net sOurs s
0 (original)1.000 ± 0.0001.000 ± 0.0001.000 ± 0.0001.000 ± 0.0001.000 ± 0.0001.000 ± 0.0001.000 ± 0.0001.000 ± 0.000
1 (0.15, 0.88)0.688 ± 0.0590.741 ± 0.0550.804 ± 0.0410.917 ± 0.0240.767 ± 0.0370.793 ± 0.0390.862 ± 0.0310.924 ± 0.021
2 (0.07, 0.76)0.607 ± 0.0530.672 ± 0.0510.754 ± 0.0420.884 ± 0.0330.709 ± 0.0330.753 ± 0.0330.819 ± 0.0320.894 ± 0.029
3 (0.03, 0.64)0.518 ± 0.0450.595 ± 0.0470.685 ± 0.0430.850 ± 0.0430.648 ± 0.0300.688 ± 0.0320.758 ± 0.0380.863 ± 0.039
4 (0.01, 0.52)0.413 ± 0.0390.483 ± 0.0440.621 ± 0.0490.811 ± 0.0560.580 ± 0.0340.620 ± 0.0360.698 ± 0.0420.827 ± 0.050
5 (0.005, 0.40)0.374 ± 0.0450.424 ± 0.0500.555 ± 0.0510.792 ± 0.0630.556 ± 0.0410.584 ± 0.0410.661 ± 0.0450.810 ± 0.057

As the reviewer points out, the variation of the features does not vanish completely. However, the degradation of our features is consistently and considerably smaller, at every level, than that of the compared methods trained with the same Dark ISP synthetic pairs. This feature-level robustness also carries over to actual detection performance, as shown by the consistent improvements across six real low-light benchmarks and three detector architectures (Table 1, Table 2, and Table R3).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.