Cross-Scenario Deraining Adaptation with Unpaired Data: Superpixel Structural Priors and Multi-Stage Pseudo-Rain Synthesis
This paper tackles the domain generalization problem in image deraining, where models trained on synthetic data fail catastrophically on out-of-distribution (OOD) real-world scenarios. The authors propose a three-stage pipeline—Superpixel Generation, Resolution-adaptive Fusion, and Pseudo-label Re-Synthesis—that adapts source-domain models to target domains using only unpaired rain-free images, eliminating the need for costly paired rainy data collection.
The paper presents a pragmatic approach to unsupervised domain adaptation for deraining by leveraging superpixel structural priors and physics-inspired rain synthesis. While the proposed modules are well-motivated and show impressive quantitative gains (up to 59% PSNR improvement), the work suffers from inconsistent experimental descriptions and an overstated originality claim. Specifically, the paper asserts it is the 'first cross-scenario deraining adaptation framework that eliminates the need for rainy target observations,' yet extensively cites prior semi-supervised and syn2real transfer learning methods (e.g., Wei et al. [41], Yasarla et al. [52]) that address similar settings with unpaired data. This weakens the contribution's novelty claim.
The superpixel-based structural transfer using SLIC is computationally efficient ($O(N)$ complexity) and the Resolution-adaptive Fusion mechanism via MSE-based texture matching provides a sensible way to inject source-domain structural priors while preserving target-domain backgrounds. The three-stage rain synthesis (salt-and-pepper noise $\rightarrow$ Gaussian blur $\rightarrow$ motion blur) is physically grounded and mimics realistic optical effects better than naïve noise addition. The plug-and-play nature demonstrated across NeRD-Rain, DRSformer, FADformer, and DFSSM validates the method's general applicability.
There are critical inconsistencies in the experimental setup: Section I and Figure 1 describe training on Rain200L and testing on Rain200H/DID/DDN as OOD domains, while Section IV-A states HQ-RAIN is the source and Rain200L is the target. This confounds reproducibility. The ablation studies are superficial—Table III only varies the fusion coefficient $\alpha$ without analyzing the sensitivity of the rain synthesis parameters ($p$, $\sigma_g$, $L_r$, $\beta$) or the superpixel compactness $m$. The claim of being 'pioneering' contradicts the cited literature on semi-supervised deraining using pseudo-labels. Furthermore, the method is only validated on synthetic-to-synthetic domain shifts, leaving open questions about transfer to real-world rainy images with complex haze, lighting variations, and perspective distortions mentioned in the introduction.
The reported PSNR gains are substantial (e.g., NeRD-Rain improves from 25.44 dB to 33.67 dB), but the experimental design raises questions about baseline fairness. The 'w/o ours' baselines appear to be direct inference without adaptation, whereas 'w/ ours' includes fine-tuning on pseudo-labeled data. A fairer comparison would involve fine-tuning with other pseudo-label generation strategies (random rain streaks, CycleGAN-based synthesis [42], or Gaussian process methods [52]). The frequency domain consistency loss $\mathcal{L}_{fft}$ and edge preservation loss are standard components, so the gains primarily stem from the data augmentation strategy rather than novel loss functions. The no-reference metrics (NIQE, PI) improvements support the perceptual quality claims, but real-world validation is absent.
The paper lacks a code availability statement or supplementary material link, which severely impedes reproducibility for a method with numerous hyperparameters. While the SLIC algorithm is standard, critical implementation details for the Resolution-adaptive Fusion (e.g., sliding window stride, random mask $M_{\text{random}}$ generation specifics) and the exact parameter ranges for rain synthesis ($L_{\min}, L_{\max}, \theta_{\min}, \theta_{\max}$) are underspecified. The training schedule mentions 200 epochs with cosine annealing, but the decay strategy for the source vs. target adaptation phases is not clarified. Without released code or exact hyperparameters, independent reproduction of the 32-59% gains would be challenging.
Image deraining plays a pivotal role in low-level computer vision, serving as a prerequisite for robust outdoor surveillance and autonomous driving systems. While deep learning paradigms have achieved remarkable success in firmly aligned settings, they often suffer from severe performance degradation when generalized to unseen Out-of-Distribution (OOD) scenarios. This failure stems primarily from the significant domain discrepancy between synthetic training datasets and the complex physical dynamics of real-world rain. To address these challenges, this paper proposes a pioneering cross-scenario deraining adaptation framework. Diverging from conventional approaches, our method obviates the requirements for paired rainy observations in the target domain, leveraging exclusively rain-free background images. We design a Superpixel Generation (Sup-Gen) module to extract stable structural priors from the source domain using Simple Linear Iterative Clustering. Subsequently, a Resolution-adaptive Fusion strategy is introduced to align these source structures with target backgrounds through texture similarity, ensuring the synthesis of diverse and realistic pseudo-data. Finally, we implement a pseudo-label re-Synthesize mechanism that employs multi-stage noise generation to simulate realistic rain streaks. This framework functions as a versatile plug-and-play module capable of seamless integration into arbitrary deraining architectures. Extensive experiments on state-of-the-art models demonstrate that our approach yields remarkable PSNR gains of up to 32% to 59% in OOD domains while significantly accelerating training convergence.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.