acceptodds
Withdrawn from ICLR 2027

BiDO: Bilevel Discriminative Learning with Online Synthesis for Object Detection in Foggy Scenes

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Object detection in foggy scenes is hampered by severe visibility degradation and the limited availability of trustworthy labels in real fog. Consequently, most methods rely on synthetic-to-real transfer, whose success depends on both the realism of synthesized fog and the detector’s ability to generalize under inevitable distribution mismatch. We present BiDO, a bilevel discriminative learning framework that performs online synthesis while learning a joint dehazing-and-detection model. BiDO couples synthesis and task learning through bilevel optimization, using task-driven signals to adversarially regulate the synthesis process during training. By introducing stronger uncertainty via online synthesis, BiDO favors flatter solutions and yields more robust generalization to real fog. We evaluate BiDO on three representative foggy-scene detection benchmarks under standard synthetic-to-real protocols and observe consistent improvements over prior state-of-the-art approaches. Additional algorithmic studies further highlight the advantages of online synthesis and demonstrate that BiDO can be integrated with different detector architectures. Code will be released if this work can be accepted.

Primary area:
learning theory
Keywords:
Object detection in foggy scenes, bilevel discriminative learning, online synthesis, joint dehazing-and-detection

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_EVcFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the problem of object detection in foggy scenes and presents the BiDO (Bilevel Discriminative Learning with Online Synthesis) framework that integrates fog synthesis, dehazing, and object detection through bilevel optimization. The fog synthesis network generates foggy images with depth-aware global synthesis and object-guided local modulation. The task model is trained for image restoration and object detection together. Results on RTTS and Foggy Driving show that the proposed framework improves performance compared to previous work.

Critical strengths
  1. BiDO achieves strong detection performance on real foggy benchmarks, with consistent improvements across nine dehazer-detector combinations in Table 4.
  2. The ablation studies show the effectiveness of the proposed components. BiDO also shows promising results with limited training samples in Table 6.
Critical weaknesses
  1. I think the justification for bilevel optimization is not sufficient. The proposed framework needs to be compared with simple alternatives, such as online adversarial augmentation [1] and curriculum-based fog training [2]. These comparisons would help clarify whether the proposed formulation provides benefits.

  2. Also, the contribution of the proposed bilevel optimization to the reported performance gains is not clear. Although BiDO substantially outperforms existing methods on RTTS (0.337 vs. 0.246 mAP), its single-level baseline already achieves 0.318 mAP. The authors should clarify which components account for the large improvement over existing methods, particularly given the differences in training protocols.

[1] X. Chen et al., "Robust and Accurate Object Detection via Adversarial Learning," CVPR 2021.

[2] C. Sakaridis et al., "Model Adaptation with Synthetic and Real Data for Semantic Dense Foggy Scene Understanding," ECCV 2018.

Questions for authors
  1. In Sections 3.2 and 5.2, the authors explain that object-guided local modulation enables more targeted fog degradation around objects. However, the global module already generates a spatially varying density field using depth information. Could the authors clarify the additional benefit of bounding-box guidance and compare it with spatially varying fog synthesis without object guidance?
  2. The authors claim in Figure 1 that bilevel optimization encourages flatter minima. Could the authors provide empirical evidence supporting this claim?
Reviewer_rUZqOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes BiDO, an adverse-weather object detection framework designed to bridge the synthetic-to-real domain gap without target-domain annotations. BiDO integrates three components: (i) a physically parameterized clear-to-fog synthesizer FπF_\pi combining global atmospheric scattering with object-guided local density perturbations Δρl\Delta \rho_l; (ii) a jointly optimized dehazing network gμg_\mu and detector HνH_\nu parameterized as θ=[μ;ν]\theta = [\mu; \nu]; and (iii) a value-function penalty first-order bilevel optimization scheme. The upper level optimizes θ\theta on synthesized images, while the lower level updates π\pi to maximize an L1L_1 discrepancy metric Φ\Phi alongside task loss Ψ\Psi. On RTTS, BiDO achieves 0.337 mAP (Faster R-CNN), outperforming a frozen clear-trained pipeline (0.318 mAP) and WARLearn (0.246 mAP). On Foggy Driving, it reports 0.215 mAP against DINO Teacher (0.208 mAP).

Critical strengths
  1. Coupling fog synthesis directly into the detector training loop via a physically parameterized model, rather than relying on static, offline augmentations, is a well-motivated approach for adverse-weather perception.
  2. The object-aware local density increment Δρl\Delta \rho_l (Eq. 2) constrains synthetic fog variations to physically plausible, depth-weighted regions, avoiding unconstrained pixel-space artifacts.
  3. Extensive Controlled Architecture Sweep: Table 4 evaluates single-level Base versus BiDO across nine architecture pairs (3 dehazers ×\times 3 detectors), showing consistent improvements (+0.010 to +0.033 mAP) across two-stage, one-stage, and DETR-style detectors.
  4. Graceful Degradation Under Data Scarcity: Table 6 demonstrates improved sample efficiency as the clear training pool shrinks from 20K to 1K images (+0.019 at 20K vs. +0.069 at 1K).
Critical weaknesses
  1. The single-level "Base" baseline trains on 20K fixed synthetic foggy images generated once offline, whereas BiDO resynthesizes fog online at every iteration. This confounds two distinct mechanisms: (a) Continuous data diversity from online sampling; (b) Task-driven bilevel feedback on π\pi. To substantiate that bilevel optimization is responsible for the gains, the paper must include an online random synthesis control (π\pi sampled from Π\Pi without task feedback) and an online Φ\Phi-only control (λ=0\lambda = 0). Without these, the observed gains cannot be attributed to task-adversarial guidance.
  2. Confounded SOTA Baselines in Tables 1 and 2
  • Table 1: The ten restoration baselines are coupled with a Faster R-CNN frozen after training on clear images. BiDO’s detector is trained jointly on synthetic fog. Comparing a fog-aware detector to a clear-trained detector evaluates task-specific fine-tuning rather than the bilevel algorithm. Notably, BiDO’s single-level Base already achieves 0.318 mAP, tying the best dehazer.
  • Table 2: Baselines (IA-YOLO, DENet, WARLearn) predominantly use YOLOv3 backbones on VOC-style inputs, whereas BiDO uses Faster R-CNN/DEA on 20K COCO images. Of the claimed 9.1 mAP margin over WARLearn, 7.2 points are already achieved by the single-level Base (0.318 vs. 0.246 mAP).
  1. Pathological Gradient Sign Structure in Eq. 13. Substituting hπ=−Φ−λΨh_\pi = -\Phi - \lambda \Psi into Eq. 13 yields the effective parameter updates:

∇θPγ=(1−γλ)∇θΨ(θt;Dπt)+γλ∇θΨ(θt;Dπ^t)\nabla_\theta P_\gamma = (1 - \gamma \lambda) \nabla_\theta \Psi(\theta_t; \mathcal{D}_{\pi_t}) + \gamma \lambda \nabla_\theta \Psi(\theta_t; \mathcal{D}_{\hat{\pi}_t})

∇πPγ=(1−γλ)∇πΨ−γ∇πΦ\nabla_\pi P_\gamma = (1 - \gamma \lambda) \nabla_\pi \Psi - \gamma \nabla_\pi \Phi

With the reported parameters γ=2.0\gamma = 2.0 and λ=0.55\lambda = 0.55, γλ=1.1\gamma \lambda = 1.1:

  • The detector update assigns a negative weight (−0.1-0.1) to samples from πt\pi_t, executing local gradient ascent on its own task loss.
  • The synthesizer update acts adversarially with a tiny net weight of γλ−1=0.1\gamma \lambda - 1 = 0.1. If γλ<1\gamma \lambda < 1, the update flips to cooperative mode (encouraging easier samples). The paper lacks sensitivity analysis around this critical threshold.
  1. The alternating value-function penalty scheme (Eqs. 9–14) is a direct instance of existing first-order penalty bilevel algorithms, particularly V-PBGD (Shen & Chen, ICML 2023), F2SA (Kwon et al., ICML 2023), and BOME (Liu et al., NeurIPS 2022). None of these foundational works are cited. The solver is presented as an original algorithmic contribution rather than an application of known penalty frameworks to vision synthesis.
  2. The core narrative that online synthesis injects uncertainty and guides the model toward flatter minima (L020–021, L176–178, L482–484) is entirely unsupported by empirical data. No Hessian spectrum, top eigenvalues, sharpness metrics (e.g., SAM-style ϵ\epsilon-sharpness), or loss landscape visualizations are provided; Figure 1 is explicitly labeled an "illustrative schematic."
  3. On Foggy Driving (101 images), BiDO’s +0.007 mAP margin (0.215 vs. 0.208) is easily within single-seed variance. Furthermore, BiDO underperforms DINO Teacher on AP50 (0.355 vs. 0.360). Methodologically, L300 states BiDO trains on "2,975 synthetic foggy images from Foggy Cityscapes", because Foggy Cityscapes is a fixed offline dataset, it remains unclear how the online learnable clear-to-fog synthesizer was applied.
  4. Every reported result is a single run with no seeds, error bars, or significance tests. Because the inner loop requires MM backpropagation steps through the dehazer, detector, and synthesizer, per-iteration training cost scales as approximately (M+2)×(M + 2)\times standard training. No training runtimes, GPU memory profiles, or compute-matched single-level baselines are provided.
Questions for authors
  1. What is the performance of online random synthesis (λ=0\lambda = 0, random π∼Π\pi \sim \Pi) and online Φ\Phi-only synthesis relative to BiDO under identical iteration budgets?
  2. Does the implementation execute gradient ascent on θ\theta with weight −0.1-0.1 as implied by γλ=1.1\gamma \lambda = 1.1? How sensitive is convergence when crossing the γλ=1.0\gamma \lambda = 1.0 boundary?
  3. How was online synthesis from clear images executed when using the pre-rendered Foggy Cityscapes dataset? Did UDA baselines receive target-domain images?
  4. Can you provide quantitative curvature measurements (e.g., top Hessian eigenvalue, loss under weight perturbations) to substantiate the "flatter minima" claim?
  5. What are the values of MM and TT, total GPU hours, and peak VRAM? How does a single-level baseline trained for equal wall-clock time perform?
Reviewer_5e3fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes an optimization strategy for object detection in foggy scenes. To address the limitations of existing methods related to their dependence on synthetic fog generation quality and the detector’s ability to generalize under distribution mismatch, it introduces a bilevel discriminative learning framework. This framework performs online fog synthesis while learning a joint dehazing and detection model, allowing the trained dehazing + detection model to achieve better performance in foggy scenes.

Critical strengths
  1. The motivation for the method is clear.
  2. The method has a reasonable basis within the dehazing + detection framework for foggy scenes. The experiments show that it can improve performance across different combinations of dehazing methods and detectors.
Critical weaknesses
  1. The experimental comparisons are not clearly described. For example, in Table 2, IA-YOLO and FAT are based on YOLO detectors, while the reported results indicate that the proposed method uses Faster R-CNN as its base detector. Do the methods in the table use the same base detector, and how does the baseline perform? The consistency of these comparisons should be checked or further clarified. In addition, the sources of the training data used by the compared methods need to be explained. For example, domain adaptation methods use unlabeled target domain images. How is a fair comparison made in this case?
  2. The method uses depth information to synthesize fog, but the compared methods do not require depth information for fog synthesis. This adds extra training cost. The paper also leaves the actual fog synthesis and dehazing operations unclear, including which parameters are updated. Without these details, it is hard to follow how the optimization is carried out.
  3. During fog synthesis, a weighted combination of dehazing and detection losses is used as feedback. How does increasing this weighted loss guarantee that the generated samples become harder for both dehazing and detection at the same time?
  4. The overall novelty of the proposed method is rather incremental, and the performance gains under the same network configurations in Table 4 are also mainly incremental improvements. The contribution mainly lies in combining fog synthesis, task feedback, joint dehazing and detection, and bilevel optimization. There is some novelty in combining these components for object detection in foggy scenes. Also, I think the main difficulty in this specific task should lie in the actual implementation of the fog synthesis constraints, rather than simply using image differences and detection losses as supervision.
Questions for authors

See the questions in the Weaknesses above, mainly the training settings of the compared methods, the source of the depth information, and the specific implementation of the fog synthesis module and physical constraints.

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

Related papers

Discussion (0)

Sign in to comment.