acceptodds
Under review as a conference paper at ICLR 2027

BiCo: Bilevel Collaborative Learning for Low-Light Vision

Abstract

Low-light conditions simultaneously impair visual quality and semantic perception, posing a fundamental challenge to robust vision in real-world environments. Rather than treating enhancement and high-level perception as isolated tasks, we propose a bilevel collaborative framework to explicitly coordinate pixel-level and semantic-level objectives for low-light vision. We first revisit the pixel-semantic gap and establish learnable role-balanced evaluators for the two levels, where pretrained foundation-model priors provide complementary quality and semantic guidance. Building upon these evaluators, we formulate the enhancer as the upper-level variable and the pixel- and semantic-level evaluators as lower-level variables, enabling reciprocal feedback between enhancement and perception through bilevel optimization. We further develop a two-stage learning strategy: cross-task collaboration learns a transferable enhancer through joint pixel-semantic optimization, while task-specific adaptation specializes the semantic network to a target downstream task with the enhancer fixed. The resulting enhancer captures task-agnostic meta-representations that generalize across different semantic tasks, including those unseen during collaborative learning. Extensive experiments on low-light detection and segmentation demonstrate consistent improvements in both visual quality and semantic performance, together with strong cross-task adaptability and generalization. The code will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Low-Light Vision, Bilevel Optimization, Image Enhancement

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.25 over 4 reviews, above 96% of the others.
Reviewer_6XapOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes BiCo, which explicitly coordinates pixel-level enhancement and semantic-level perception under low light via bilevel optimization. It designs two role-balanced evaluators that provide complementary quality and semantic feedback, optimizes the enhancer at the upper level while alternating the two evaluators at the lower level, and adopts a two-stage scheme of cross-task collaboration plus task-specific adaptation.

Critical strengths

The bidirectional bilevel modeling is a substantive improvement over one-way cascading. The hierarchical dependency between the upper-level enhancer and the two lower-level evaluators is explicitly modeled, and V-PBGD avoids differentiating through the alternating trajectory, achieving mutual feedback between pixel and semantic objectives with tractable computation.

Critical weaknesses
  1. Only two downstream tasks are covered. The cross-task conclusions of CTC and TSA are validated only between detection and segmentation, with no extrapolation to keypoints, depth estimation, or video tasks, leaving the boundary of transferability unclear.
  2. A direct comparison with the Lipschitz-view restoration-detection synergy work of Zhao et al. 2025 is missing, although it is the closest in motivation and is only mentioned in related work without experimental comparison.
  3. Enhancement quality evidence is limited and not fully independent. Visual quality relies only on no-reference IQA metrics, which give inconsistent rankings, and CLIP-IQA overlaps with the CLIP-based guidance used in training. Paired-reference metrics or a human study would better support the claim.
Questions for authors
  1. Can you provide reference-based PSNR, SSIM, and LPIPS to verify pixel-level fidelity of the enhancement?
  2. Do the cross-task conclusions generalize beyond detection and segmentation, for example to keypoint estimation or video understanding?
  3. How does BiCo compare numerically with the Lipschitz-based synergy method of Zhao et al. 2025 under the same settings?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 6Xap

Table 1: Quantitative comparison of full-reference image quality on the MIT dataset.

MethodPSNR ↑SSIM ↑LPIPS ↓
ZeroDCE15.290.7440.184
RUAS12.280.6930.223
SCI14.880.7500.206
PairLIE12.900.6610.263
LightenDiff19.620.789‾\underline{0.789}0.175
ZERO-IG14.120.6690.244
QuadPrior17.410.7590.180
Di-Retinex11.220.6810.216
CWNet17.960.7820.153‾\underline{0.153}
BiCo (Ours)18.84‾\underline{18.84}0.8210.140

Q1: Can you provide reference-based PSNR, SSIM, and LPIPS to verify pixel-level fidelity of the enhancement?

Response: Our main focus is joint enhancement-perception learning, where the enhancer is trained without paired reference images, reducing dependence on supervised pixel-level reconstruction and favoring broader downstream applicability. Nevertheless, following the reviewer's suggestion, we additionally evaluate full-reference fidelity on the standard MIT test set, consisting of 500 paired samples [1], using PSNR, SSIM, and LPIPS.

As shown in Table 1, BiCo achieves the best SSIM and LPIPS and the second-best PSNR among the compared methods. These results demonstrate that, despite being optimized without paired supervision, BiCo preserves strong pixel-level fidelity while simultaneously supporting downstream semantic perception.

[1] Ma L, Ma T, Xu C, et al. Learning with self-calibrator for fast and robust low-light image enhancement[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025.

Table 2: Quantitative comparison of joint image enhancement–instance segmentation (En + In) and joint image enhancement–depth estimation (En + Dep).

MethodEn‾\underline{\mathrm{En}} + In‾\underline{\mathrm{In}}En‾\underline{\mathrm{En}} + In‾\underline{\mathrm{In}}En‾\underline{\mathrm{En}} + In‾\underline{\mathrm{In}}En‾\underline{\mathrm{En}} + In‾\underline{\mathrm{In}}En‾\underline{\mathrm{En}} + Dep‾\underline{\mathrm{Dep}}En‾\underline{\mathrm{En}} + Dep‾\underline{\mathrm{Dep}}En‾\underline{\mathrm{En}} + Dep‾\underline{\mathrm{Dep}}En‾\underline{\mathrm{En}} + Dep‾\underline{\mathrm{Dep}}
LIQE ↑CLIPIQA ↑HyperIQA ↑mAP (%) ↑LIQE ↑CLIPIQA ↑HyperIQA ↑Abs Rel ↓
ZeroDCE1.2220.2550.30754.691.0100.1510.2350.322
SCI1.4510.2690.31954.441.0040.1530.2170.315‾\underline{0.315}
LightenDiff1.2850.291‾\underline{0.291}0.31854.011.011‾\underline{1.011}0.1220.239‾\underline{0.239}0.316
Di-Retinex1.1000.1840.28954.561.0020.1950.2380.354
CWNet1.3100.2180.36054.83‾\underline{54.83}1.0020.1070.2130.339
BiCo (Ours)1.439‾\underline{1.439}0.3180.326‾\underline{0.326}54.881.0220.156‾\underline{0.156}0.3350.310

Q2: Do the cross-task conclusions generalize beyond detection and segmentation?

Response: We further evaluate BiCo on two additional downstream tasks. For comparison, we select the five top-performing competing enhancers based on their average detection and segmentation performance in Table 2 of the manuscript. Specifically, instance segmentation is evaluated on the RGB-Dark split of LIS [1] using RF-DETR-Seg Small [2], while monocular depth estimation is evaluated on nuScenes-Night [3] using Depth-Anything-V2-Metric-VKITTI-Base [4]. For all methods, the enhancers are kept fixed, and only the downstream task networks are fine-tuned. As shown in Table 2, BiCo achieves the best downstream performance on both tasks while maintaining competitive enhancement quality, further demonstrating the transferability of its learned meta-representations across diverse downstream tasks.

[1] Chen, et al., "Instance Segmentation in the Dark," IJCV 2023.

[2] Robinson, et al., "RF-DETR: Neural Architecture Search for Real-Time Detection Transformers," ICLR 2026.

[3] Zheng, et al., "STEPS: Joint Self-supervised Nighttime Image Enhancement and Depth Estimation," ICRA 2023.

[4] Yang, et al., "Depth Anything V2," NeurIPS 2024.

Table 3: Quantitative comparison with the Lipschitz-based restoration-detection synergy method of Zhao et al. on ExDark.

MethodLIQE ↑CLIPIQA ↑HyperIQA ↑mAP ↑
LR-YOLO1.4940.2960.36530.15
BiCo (Ours)2.0080.4370.43143.55

Q3: How does BiCo compare numerically with the Lipschitz-based synergy method of Zhao et al. 2025 under the same settings?

Response: We additionally compare BiCo with Lipschitz-based synergy method LR-YOLO under the same ExDark setting. As shown in Table 3, BiCo consistently outperforms LR-YOLO in both enhancement quality and detection performance.

Dear Reviewer 6Xap,

Could you check if the response has addressed your issues? Since the discussion is closing soon, your reply would be very important.

Best,

AC

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

This paper investigates how low-light image enhancement and downstream semantic perception can be learned collaboratively. The authors propose BiCo, a bilevel framework that treats the enhancer as the upper-level variable and VLM-guided pixel-level and VFM-guided semantic-level evaluators as lower-level components. Training consists of cross-task collaborative learning followed by task-specific adaptation with the enhancer frozen. The evaluation covers low-light object detection and semantic segmentation, with ablation studies and cross-task transfer experiments examining the contributions of the framework and the transferability of the learned enhancer.

Critical strengths
  1. The paper revisits low-light vision from both forward and backward perspectives: enhancement provides visual evidence for perception, while semantic supervision guides enhancement through backpropagation. This perspective offers a useful way to examine the limitations of conventional enhancement–perception pipelines and clearly motivates collaboration between the two objectives.

  2. The proposed bilevel framework seeks to establish a collaborative relationship between image enhancement and semantic perception. Compared with conventional joint optimization, it explicitly formulates their hierarchical dependency, providing a coherent approach to coordinating pixel-level and semantic-level objectives.

  3. The paper reports competitive results on object detection and semantic segmentation, supported by ablation studies and cross-task transfer analysis. In particular, it improves both visual quality and downstream performance without increasing the complexity of the enhancement architecture, highlighting the practical potential of the proposed training strategy.

Critical weaknesses
  1. A sensitivity analysis of key hyperparameters would complement the ablation studies in Table 6. In particular, varying the regularization weight and reporting both image quality and downstream accuracy would help clarify the balance between task supervision and semantic regularization, as well as the robustness of the chosen setting.

  2. Other issues: (1) The “CTC” rows in Table 8 report 43.55 mAP and 63.23 mIoU, which correspond to “CTC + TSA” in Table 7. Please clarify which checkpoints are used for model recombination and make the stage labels consistent. (2) Section 4.3 refers to Figure 4 when discussing RIR’s misclassification of a dog as a cup, whereas the RIR examples appear in Figure 3. The figure references and captions should be checked for consistency. (3) In Section 3.1, two consecutive sentences begin with “As shown in Table 1.” Combining or rephrasing these sentences would improve readability.

Questions for authors
  1. Could the authors clarify how the shared parameters are updated and how the value functions and their gradients are approximated in practice? How do these steps correspond to the proposed bilevel formulation?

  2. How do the authors explain the observation in Table 6 that the semantic-only constraint achieves higher image-quality scores than the pixel-only constraint? What complementary contribution does the pixel-level constraint provide when combined with semantic guidance?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Lejg (Part 1)

Q1: A sensitivity analysis of key hyperparameters would complement the ablation studies.

*Response: In BiCo, Lp\mathcal{L}_p and Ls\mathcal{L}_s serve as the two lower-level constraints for pixel-level and semantic-level evaluation, respectively, whose contributions are examined in the constraint ablation. For hyperparameter sensitivity, we further investigate λr\lambda_r, which balances the VFM-guided regularization term Lr\mathcal{L}_r in Ls\mathcal{L}_s. As shown in Table 1, λr=0.05\lambda_r=0.05 achieves the best overall performance, while the results remain relatively stable across the tested values.

Table 1: Sensitivity analysis of the semantic regularization weight λr\lambda_r on joint enhancement-detection.

λr\lambda_rLIQE ↑CLIPIQA ↑HyperIQA ↑mAP ↑
0.0251.9970.4340.42643.37
0.052.0100.4370.43143.55
0.0751.9820.4310.43143.33

Q2: Other issues: (1) The “CTC” rows in Table 8 report 43.55 mAP and 63.23 mIoU, which correspond to “CTC + TSA” in Table 7. Please clarify which checkpoints are used for model recombination and make the stage labels consistent. (2) Section 4.3 refers to Figure 4 when discussing RIR’s misclassification of a dog as a cup, whereas the RIR examples appear in Figure 3. The figure references and captions should be checked for consistency. (3) In Section 3.1, two consecutive sentences begin with “As shown in Table 1.” Combining or rephrasing these sentences would improve readability.

Response: We thank the reviewer for pointing out these issues. We have carefully checked and revised the manuscript as follows:

(1) We apologize for the incorrect stage labeling in Table 8. The reported results (43.55 mAP and 63.23 mIoU) correspond to the complete CTC+TSA pipeline, rather than CTC alone. We have corrected the stage labels accordingly and clarified that model recombination uses the enhancer checkpoints learned during CTC together with the corresponding downstream task checkpoints fine-tuned during TSA.

(2) We apologize for the inconsistent figure references. The figure order was inadvertently reversed. We have corrected the figure ordering and revised the corresponding descriptions and captions to ensure consistency throughout Section 4.3.

(3) We have rephrased the two consecutive sentences beginning with "As shown in Table 1" and further refined the surrounding discussion to improve readability and logical flow.

All corresponding corrections have been incorporated into the revised manuscript.

Q3: Could the authors clarify how the shared parameters are updated and how the value functions and their gradients are approximated in practice? How do these steps correspond to the proposed bilevel formulation?

Response: We decompose the lower-level parameters as π=ϑ∪κ{\pi}={\vartheta}\cup{\kappa}, where ϑ{\vartheta} denotes the shared evaluator parameters and κ{\kappa} denotes the semantic-specific parameters. As formulated in Eq. (7), ϑ{\vartheta} is updated using the pixel-level loss Lp\mathcal{L}_p, while κ{\kappa} is updated using the semantic-level loss Ls\mathcal{L}_s with the updated ϑ{\vartheta}. We set K=2K=2 inner iterations to approximate the lower-level minima and corresponding value functions. Their approximations are incorporated into the penalized objective in Eq. (10), whose gradient is used to update the upper-level enhancer θ{\theta} via projected gradient descent (Eq. (11)), without differentiating through the full lower-level optimization trajectory. This follows the V-PBGD framework, which provides finite-time convergence guarantees under appropriate conditions.

AuthorsOpenReview ↗

Response to Reviewer Lejg (Part 2)

Q4: How do the authors explain the observation in Table 6 that the semantic-only constraint achieves higher image-quality scores than the pixel-only constraint? What complementary contribution does the pixel-level constraint provide when combined with semantic guidance?

Response: Following another reviewer's suggestion, we have further expanded this set of experiments and updated the corresponding results. Specifically, Table 2 compares the Naive baseline, two single-level optimization strategies (Joint and Alternating), and our bilevel approach (BiCo). For Joint, Alternating, and BiCo, we evaluate four configurations with neither, only pixel-level (Lp\mathcal{L}_p), only semantic-level (Ls\mathcal{L}_s), or both constraints enabled. We additionally report enhancement quality (LIQE, CLIPIQA, and HyperIQA), downstream detection performance (mAP), training time, memory consumption, and inference parameters.

An important finding from Table 2 is that, under both Joint and BiCo, using only semantic-level guidance consistently achieves better enhancement quality and detection performance than using only pixel-level guidance. This observation is particularly meaningful because our pixel-level constraint is also derived from VLM/VFM semantic priors, rather than conventional pixel-level supervision. However, it primarily exploits these priors to guide pixel-level enhancement, whereas the semantic-level constraint explicitly preserves and aligns high-level semantic representations. The consistent improvements across all three enhancement-quality metrics indicate that semantic-level guidance provides effective structural and semantic support to the enhancement process itself, rather than merely benefiting downstream detection.

Furthermore, combining pixel-level and semantic-level constraints yields the best detection performance with BiCo, demonstrating their complementary roles in visual quality preservation and task-relevant semantic representation learning. Although BiCo incurs additional training costs due to bilevel optimization, all methods maintain the same inference parameter count. These additional experiments and the corresponding analysis are included in the revised manuscript.

Table 2: Comparison of single-level and bilevel optimization under different pixel-semantic constraints on the detection task.

OptimizationStrategyLp\mathcal{L}_pLs\mathcal{L}_sLIQE ↑CLIPIQA ↑HyperIQA ↑mAP ↑Time (H) ↓Memory (GB) ↓Params (M) ↓
Single-LevelNaive××1.7850.4320.41542.154.4053.91132.044
Single-LevelJoint××1.9560.4300.42642.385.0145.98732.044
Single-LevelJoint✓×1.9460.4300.42842.455.12412.2832.044
Single-LevelJoint×✓2.0200.4330.431‾\underline{0.431}42.914.8017.28432.044
Single-LevelJoint✓✓2.0040.4340.42743.125.81813.5432.044
Single-LevelAlternating××1.9890.4330.431‾\underline{0.431}38.725.6923.89232.044
Single-LevelAlternating✓×1.9920.4340.43039.755.94011.3932.044
Single-LevelAlternating×✓1.9910.4330.42840.506.4265.10632.044
Single-LevelAlternating✓✓1.9970.4340.431‾\underline{0.431}40.116.74212.6632.044
BilevelBiCo (Ours)××1.9850.4340.43242.4410.1711.1132.044
BilevelBiCo (Ours)✓×1.9610.436‾\underline{0.436}0.42742.7311.6314.6232.044
BilevelBiCo (Ours)×✓2.010‾\underline{2.010}0.4370.431‾\underline{0.431}43.05‾\underline{43.05}10.9612.4732.044
BilevelBiCo (Ours)✓✓2.0080.4370.431‾\underline{0.431}43.5514.2816.0832.044

Thank you for the response, which addresses my questions.

The added sensitivity analysis and single-/bilevel comparisons strengthen the paper, and I believe this work is meaningful for low-light vision. I have raised my score. Please include these results and clarifications in the revised version.

AuthorsOpenReview ↗

Response to Reviewer Lejg

Dear Reviewer,

Thank you for your positive feedback and recognition of our work. We appreciate your constructive comments and are glad that the additional analyses and clarifications have strengthened the paper. We have incorporated the suggested revisions and uploaded the updated version of the manuscript for your consideration.

Reviewer_aE8ROpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents BiCo, a bilevel collaborative learning framework for low-light vision that jointly considers image enhancement and downstream semantic perception. The authors argue that conventional cascaded approaches either optimize enhancement without semantic awareness or allow semantic supervision to dominate the enhancer, creating a trade-off between visual fidelity and task performance.

BiCo treats the enhancer as an upper-level variable and two coupled evaluators as lower-level variables. The pixel-level evaluator combines learnable features with frozen CLIP-based quality cues, while the semantic-level evaluator augments the task loss with bidirectional spatial contrastive alignment to frozen DINO features. The resulting problem is optimized using a value-function-based bilevel strategy with two stages: cross-task collaboration learns the enhancer, and task-specific adaptation freezes the enhancer while fine-tuning the downstream model.

Experiments are conducted on ExDark object detection and ACDC semantic segmentation. BiCo is compared with cascaded enhancement pipelines and end-to-end low-light perception methods, with the reported results showing improvements in no-reference image-quality metrics, mAP, and mIoU.

Critical strengths

A well-motivated bilevel formulation of pixel–semantic collaboration. The paper identifies a meaningful limitation of conventional low-light pipelines: pixel-oriented enhancement may discard task-relevant information, while direct semantic supervision may dominate the enhancer and harm visual fidelity. Modeling the enhancer as an upper-level variable and the pixel/semantic evaluators as coupled lower-level variables provides a principled alternative to a simple weighted multi-task objective.

Critical weaknesses

The experiments do not cleanly isolate the contribution of bilevel optimization. The “Naive (Single Level)” baseline is insufficiently specified, and its reported values exactly match the SCI baseline, making it unclear whether it uses the same VLM/VFM evaluators, losses, update schedule, and computational budget as BiCo. The paper also lacks comparisons against joint weighted optimization and computation-matched alternating training. Consequently, the reported gains cannot be confidently attributed to the paper’s central bilevel formulation rather than to the added foundation-model objectives or additional optimization steps.

Questions for authors

What exactly is the “Naive (Single Level)” baseline? Does it use the same evaluators, losses, training schedule, and computational budget as BiCo?

How are the lower-level optima and value functions approximated in practice, and how many inner-loop updates are used?

How much additional training time, memory, and computation does BiCo require compared with the single-level baseline?

Are the reported results averaged over multiple random seeds? Please provide variance estimates, particularly for the small cross-task improvements.

Given that transfer is evaluated on only two tasks and one direction yields only a minor gain with degraded quality metrics, what evidence supports the stronger claim of task-agnostic meta-representations?

3 comments
AuthorsOpenReview ↗

Response to Reviewer aE8R (Part 1)

Q1: What exactly is the “Naive (Single Level)” baseline? Does it use the same evaluators, losses, training schedule, and computational budget as BiCo?

Response: The “Naive” baseline follows a conventional one-way enhancement-to-perception pipeline, where enhanced images are fed into the downstream network, and only the downstream network is fine-tuned, without updating the enhancer or introducing pixel-semantic collaboration. To provide more comprehensive comparisons and isolate the benefit of bilevel optimization, we further introduce Joint and Alternating single-level strategies under the same combinations of Lp\mathcal{L}_p and Ls\mathcal{L}_s as BiCo. We also report training time, memory, and inference parameters for a comprehensive comparison. As shown in Table 1, although BiCo incurs higher training costs due to bilevel optimization and additional lower-level updates, it consistently achieves higher mAP than both single-level strategies under the corresponding guidance settings. Importantly, all methods retain identical inference parameter counts, confirming that these gains do not require additional inference capacity.

Table 1: Comparison of single-level and bilevel optimization under different pixel-semantic constraints on the detection task.

OptimizationStrategyLp\mathcal{L}_pLs\mathcal{L}_sLIQE ↑CLIPIQA ↑HyperIQA ↑mAP ↑Time (H) ↓Memory (GB) ↓Params (M) ↓
Single-LevelNaive××1.7850.4320.41542.154.4053.91132.044
Single-LevelJoint××1.9560.4300.42642.385.0145.98732.044
Single-LevelJoint✓×1.9460.4300.42842.455.12412.2832.044
Single-LevelJoint×✓2.0200.4330.431‾\underline{0.431}42.914.8017.28432.044
Single-LevelJoint✓✓2.0040.4340.42743.125.81813.5432.044
Single-LevelAlternating××1.9890.4330.431‾\underline{0.431}38.725.6923.89232.044
Single-LevelAlternating✓×1.9920.4340.43039.755.94011.3932.044
Single-LevelAlternating×✓1.9910.4330.42840.506.4265.10632.044
Single-LevelAlternating✓✓1.9970.4340.431‾\underline{0.431}40.116.74212.6632.044
BilevelBiCo (Ours)××1.9850.4340.43242.4410.1711.1132.044
BilevelBiCo (Ours)✓×1.9610.436‾\underline{0.436}0.42742.7311.6314.6232.044
BilevelBiCo (Ours)×✓2.010‾\underline{2.010}0.4370.431‾\underline{0.431}43.05‾\underline{43.05}10.9612.4732.044
BilevelBiCo (Ours)✓✓2.0080.4370.431‾\underline{0.431}43.5514.2816.0832.044

Q2: How are the lower-level optima and value functions approximated in practice, and how many inner-loop updates are used?

Response: We adopt the value-function-based penalty bilevel gradient descent (V-PBGD) method, rather than developing a new solver. The lower-level value functions are approximated using K=2K=2 inner iterations, and their values are incorporated into the penalized objective. The upper-level gradient is then computed from this objective to update the enhancer via projected gradient descent. Importantly, V-PBGD establishes finite-time convergence guarantees for this approach under appropriate problem assumptions.

Q3: How much additional training time, memory, and computation does BiCo require compared with the single-level baseline?

Response: As shown in Table 1, BiCo incurs higher training time and memory costs due to the additional computations required by bilevel optimization. Nevertheless, these costs are limited to training, while BiCo achieves clear performance gains without increasing inference complexity.

AuthorsOpenReview ↗

Response to Reviewer aE8R (Part 2)

Q4: Are the reported results averaged over multiple random seeds? Please provide variance estimates, particularly for the small cross-task improvements.

Response: The original results were obtained using a single seed. To assess their stability, we additionally evaluate BiCo with three random seeds and report the mean and standard deviation in Table 2. Notably, the cross-task results exhibit small variations, with standard deviations of 0.08 mIoU for detection-to-segmentation transfer and 0.05 mAP for segmentation-to-detection transfer, confirming the consistency of the observed gains.

Table 2: Training stability of BiCo across different random seeds on two joint tasks.

CTC SettingRandom SeedTSA: En‾\underline{\mathrm{En}} + De‾\underline{\mathrm{De}} (mAP ↑)TSA: En‾\underline{\mathrm{En}} + Se‾\underline{\mathrm{Se}} (mIoU ↑)
En‾\underline{\mathrm{En}} + De‾\underline{\mathrm{De}}Seed 143.5562.69
En‾\underline{\mathrm{En}} + De‾\underline{\mathrm{De}}Seed 243.5662.56
En‾\underline{\mathrm{En}} + De‾\underline{\mathrm{De}}Seed 343.5362.70
En‾\underline{\mathrm{En}} + De‾\underline{\mathrm{De}}Mean ±\pm Std43.55 ±\pm 0.0262.65 ±\pm 0.08
En‾\underline{\mathrm{En}} + Se‾\underline{\mathrm{Se}}Seed 142.3663.23
En‾\underline{\mathrm{En}} + Se‾\underline{\mathrm{Se}}Seed 242.3863.17
En‾\underline{\mathrm{En}} + Se‾\underline{\mathrm{Se}}Seed 342.2963.24
En‾\underline{\mathrm{En}} + Se‾\underline{\mathrm{Se}}Mean ±\pm Std42.34 ±\pm 0.0563.21 ±\pm 0.04

Q5: Given that transfer is evaluated on only two tasks and one direction yields only a minor gain with degraded quality metrics, what evidence supports the stronger claim of task-agnostic meta-representations?

Response: We clarify that the meta-representation claim specifically refers to the enhancer learned during Cross-Task Collaboration (CTC). To further examine its transferability beyond detection and semantic segmentation, we introduce two additional downstream tasks: low-light instance segmentation and monocular depth estimation. We select the five top-performing competing enhancers based on their average detection and segmentation performance in Table 2 (manuscript). Specifically, instance segmentation is evaluated on the RGB-Dark split of LIS [1] using RF-DETR-Seg Small [2], while monocular depth estimation is evaluated on nuScenes-Night [3] using Depth-Anything-V2-Metric-VKITTI-Base [4]. For all methods, the enhancers are kept fixed, and only the downstream task networks are fine-tuned. As shown in Table 3, BiCo achieves the best downstream performance on both tasks while maintaining competitive enhancement quality. Together with the cross-task recombination results in Table 8, these findings further support that the CTC-learned enhancer captures transferable meta-representations rather than being tightly coupled to specific semantic tasks.

[1] Chen, et al., "Instance Segmentation in the Dark," IJCV 2023.

[2] Robinson, et al., "RF-DETR: Neural Architecture Search for Real-Time Detection Transformers," ICLR 2026.

[3] Zheng, et al., "STEPS: Joint Self-supervised Nighttime Image Enhancement and Depth Estimation," ICRA 2023.

[4] Yang, et al., "Depth Anything V2," NeurIPS 2024.

Table 3: Quantitative comparison of joint image enhancement–instance segmentation (En + In) and joint image enhancement–depth estimation (En + Dep).

MethodEn‾\underline{\mathrm{En}} + In‾\underline{\mathrm{In}}En‾\underline{\mathrm{En}} + In‾\underline{\mathrm{In}}En‾\underline{\mathrm{En}} + In‾\underline{\mathrm{In}}En‾\underline{\mathrm{En}} + In‾\underline{\mathrm{In}}En‾\underline{\mathrm{En}} + Dep‾\underline{\mathrm{Dep}}En‾\underline{\mathrm{En}} + Dep‾\underline{\mathrm{Dep}}En‾\underline{\mathrm{En}} + Dep‾\underline{\mathrm{Dep}}En‾\underline{\mathrm{En}} + Dep‾\underline{\mathrm{Dep}}
LIQE ↑CLIPIQA ↑HyperIQA ↑mAP (%) ↑LIQE ↑CLIPIQA ↑HyperIQA ↑Abs Rel ↓
ZeroDCE1.2220.2550.30754.691.0100.1510.2350.322
SCI1.4510.2690.31954.441.0040.1530.2170.315‾\underline{0.315}
LightenDiff1.2850.291‾\underline{0.291}0.31854.011.011‾\underline{1.011}0.1220.239‾\underline{0.239}0.316
Di-Retinex1.1000.1840.28954.561.0020.1950.2380.354
CWNet1.3100.2180.36054.83‾\underline{54.83}1.0020.1070.2130.339
BiCo (Ours)1.439‾\underline{1.439}0.3180.326‾\underline{0.326}54.881.0220.156‾\underline{0.156}0.3350.310

Dear Reviewer aE8R,

Could you check if the response has addressed your issues? Since the discussion is closing soon, your reply would be very important.

Best,

AC

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

BiCo presents a bilevel collaborative framework for low-light enhancement and semantic perception. It combines learnable pixel and semantic evaluators with foundation-model guidance, followed by task-specific adaptation with the enhancer frozen. Its main contribution is an optimization strategy for coordinating visual quality and downstream performance.

Critical strengths
  1. The paper addresses a meaningful mismatch between visual enhancement and semantic utility.
  2. A lightweight enhancer benefits detection and segmentation, with ablations supporting complementary pixel/semantic guidance and downstream adaptation.
  3. Cross-task transfer and enhancer-recombination experiments provide useful evidence beyond task-matched evaluation.
Critical weaknesses
  1. Retinexformer and GLARE report ExDark detection experiments and should be evaluated under the same downstream protocol. The manuscript should also specify whether enhancement baselines use official checkpoints or retrained models, and whether downstream training and adaptation budgets are matched.

  2. Unclear optimization details. The approximations to the value-function minima and upper-level gradients in Eq. (10) are insufficiently explained. The relationship between the shared-parameter formulation in Eq. (5) and the blockwise updates in Eqs. (7)–(9) also requires clarification.

  3. Modest gains and limited transfer evidence. Detection improves from 42.98 to 43.55 mAP and segmentation from 62.91 to 63.23 mIoU over unenhanced inputs. Transfer is evaluated only between these two tasks, with segmentation-to-detection gaining 0.21 mAP over the naive baseline while reducing some quality scores. The “task-agnostic meta-representation” claim should be qualified.

Questions for authors

My current assessment is borderline and is based in part on the assumption that the authors can satisfactorily address most of these concerns. I hope the response can provide concrete explanations and supporting evidence, particularly on the following points:

  1. Why are Retinexformer and GLARE omitted, and what checkpoint, training, and downstream-adaptation protocols are used for all enhancement baselines?
  2. Does BiCo retain its advantage over single-level joint training with identical losses, foundation-model guidance, and total training budgets?
  3. How are the value-function minima and upper-level gradients approximated, and are the reported gains stable across training seeds?
3 comments
AuthorsOpenReview ↗

Response to Reviewer cQfF (Part 1)

Q1: Why are Retinexformer and GLARE omitted?

Response: Our main focus is joint enhancement-perception learning. Accordingly, in the related work, we primarily discuss reference-free enhancement paradigms, and our enhancement-based comparisons also mainly follow this setting. Such methods rely less on paired supervision and are generally more favorable for downstream generalization than supervised methods such as Retinexformer and GLARE. Following the reviewer’s suggestion, we additionally include both methods in Table 1. BiCo remains competitive, achieving the best CLIPIQA and the highest detection mAP.

Table 1: Quantitative comparison with Retinexformer and GLARE.

MethodLIQE ↑CLIPIQA ↑HyperIQA ↑mAP ↑
Retinexformer1.7550.3060.39241.30
GLARE2.2020.397‾\underline{0.397}0.45141.72
BiCo (Ours)2.008‾\underline{2.008}0.4370.431‾\underline{0.431}43.55

Q2: What checkpoint, training, and downstream-adaptation protocols are used for all enhancement baselines?

Response: All enhancement baselines use their officially released pretrained checkpoints, with the enhancers frozen and only the downstream networks fine-tuned. Both our TSA stage and baseline fine-tuning follow the same standard MMDetection and MMSegmentation configurations. Specifically, detection uses SGD (learning rate: 1.25×10−31.25 \times 10^{-3}, momentum: 0.9, weight decay: 1×10−41 \times 10^{-4}), while segmentation uses AdamW (learning rate: 1.5×10−51.5 \times 10^{-5}, weight decay: 1×10−21 \times 10^{-2}). For a fair comparison, all baselines are trained for 29 epochs (detection) or 88k iterations (segmentation), matching the total training budgets of our CTC and TSA stages (24+5 epochs and 80k+8k iterations, respectively). The final checkpoints are used for evaluation.

Q3: Does BiCo retain its advantage over single-level joint training with identical losses, foundation-model guidance, and total training budgets?

Response: To isolate the benefit of bilevel optimization, we compare two single-level strategies: joint optimization, which simultaneously optimizes enhancement and task-specific objectives, and alternating optimization, which updates them alternately. We further evaluate all strategies under identical combinations of Lp\mathcal{L}_p and Ls\mathcal{L}_s to disentangle the effects of VLM/VFM guidance.

As shown in Table 2, under the same full pixel-semantic guidance, BiCo achieves the highest mAP and consistently improves all three enhancement-quality metrics over the single-level joint counterpart. The increased training cost primarily stems from nested optimization and additional lower-level updates required by bilevel learning. Importantly, this overhead is confined to training, with all variants maintaining the same inference parameter count.

Table 2: Comparison of single-level and bilevel optimization under different pixel-semantic constraints on the detection task.

OptimizationStrategyLp\mathcal{L}_pLs\mathcal{L}_sLIQE ↑CLIPIQA ↑HyperIQA ↑mAP ↑Time (H) ↓Memory (GB) ↓Params (M) ↓
Single-LevelNaive××1.7850.4320.41542.154.4053.91132.044
Single-LevelJoint××1.9560.4300.42642.385.0145.98732.044
Single-LevelJoint✓×1.9460.4300.42842.455.12412.2832.044
Single-LevelJoint×✓2.0200.4330.431‾\underline{0.431}42.914.8017.28432.044
Single-LevelJoint✓✓2.0040.4340.42743.125.81813.5432.044
Single-LevelAlternating××1.9890.4330.431‾\underline{0.431}38.725.6923.89232.044
Single-LevelAlternating✓×1.9920.4340.43039.755.94011.3932.044
Single-LevelAlternating×✓1.9910.4330.42840.506.4265.10632.044
Single-LevelAlternating✓✓1.9970.4340.431‾\underline{0.431}40.116.74212.6632.044
BilevelBiCo (Ours)××1.9850.4340.43242.4410.1711.1132.044
BilevelBiCo (Ours)✓×1.9610.436‾\underline{0.436}0.42742.7311.6314.6232.044
BilevelBiCo (Ours)×✓2.010‾\underline{2.010}0.4370.431‾\underline{0.431}43.05‾\underline{43.05}10.9612.4732.044
BilevelBiCo (Ours)✓✓2.0080.4370.431‾\underline{0.431}43.5514.2816.0832.044
AuthorsOpenReview ↗

Response to Reviewer cQfF (Part 2)

Q4: How are the value-function minima and upper-level gradients approximated, and are the reported gains stable across training seeds?

Response: Our focus is not on designing a new bilevel solver, but on how to model and understand the interaction between pixel-level enhancement and semantic perception through bilevel optimization. For optimization, we directly adopt the penalty-based bilevel gradient descent method in [1], where the lower-level minima are approximated iteratively and the upper-level variable is updated through the penalized objective. The method in [1] provides finite-time convergence guarantees under appropriate conditions.

Table 3: Training stability of BiCo across different random seeds.

Random SeedEn‾\underline{\mathrm{En}} + De‾\underline{\mathrm{De}} (mAP ↑)En‾\underline{\mathrm{En}} + Se‾\underline{\mathrm{Se}} (mIoU ↑)
Seed 143.5563.23
Seed 243.5663.17
Seed 343.5363.24
Mean ± Std43.55 ± 0.0263.21 ± 0.04

We further evaluate BiCo across three random seeds (Table 3), observing consistently low variations in both detection and segmentation performance, with standard deviations of only 0.02 mAP and 0.04 mIoU, respectively.

[1] H. Shen and T. Chen, “On Penalty-based Bilevel Gradient Descent Method,” ICML, 2023.

I thank the authors for their response.

Regarding Q1 and Q2, I am satisfied with the current rebuttal. However, I strongly recommend that the authors discuss the LLIE baselines that report detection results on ExDark in the final version, even though these methods are supervised. More importantly, the claim that “reference-free methods are generally more favorable for downstream generalization than supervised methods” does not always hold, as evidenced by GLARE’s performance being comparable to that of BiCo in Table 1. In addition, some baselines may provide multiple official checkpoints trained on different datasets, such as LOLv1 and LOLv2-Real. Please clearly specify which checkpoints were used in the revised version.

Overall, given the meaningful problem formulation and potential practical applications, I will maintain my positive rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.