Weather-Aware Domain Adaptation for Street-View Weather Recognition
This work tackles street-view weather recognition under severe domain shift, where labeled training images come from generic outdoor datasets and real driving scenes are unlabeled at training time. The proposed Weather-Aware ADDA conditions the adversarial domain discriminator on predicted weather probabilities, encouraging alignment that remains sensitive to weather semantics. The idea is to enable reliable onboard camera-based perception for autonomous vehicles without requiring expensive labeled street-view data for every adverse condition.
WA-ADDA is a sensible, application-driven extension of ADDA for weather classification, but its contribution is more incremental than the authors frame. The architectural change is limited to concatenating predicted class probabilities $p$ with features $z$ as input to the domain discriminator $D_\phi([z \parallel p])$, a form of auxiliary conditioning that has precedent in adversarial adaptation literature. More critically, the proposed method bundles this discriminator change with entropy minimization and consistency regularization, yet compares the full bundle against vanilla ADDA without either addition, so the gain cannot be cleanly attributed to weather-aware conditioning. The benchmark is useful, but baseline coverage is thin—only Source-Only and ADDA are included, omitting standard alternatives such as DANN, CORAL, or MMD-based methods.
The core intuition—that domain alignment should respect weather semantics—is well-motivated for autonomous driving perception. The multi-source benchmark and emphasis on macro accuracy are practical contributions for an imbalanced problem, and the cross-backbone evaluation shows the gains are not architecture-specific. The t-SNE visualizations lend qualitative support: "WA-ADDA produces more coherent and semantically meaningful feature distributions" and target embeddings integrate better into the source space than under vanilla ADDA. Additionally, the standardized protocol with per-class recalls is a step forward for reproducibility in weather recognition.
The strongest concern is the lack of isolation between the discriminator conditioning and the auxiliary target-side losses. As the authors note, WA-ADDA "extends the standard ADDA framework with two key modifications: (i) the discriminator jointly receives both feature-level embeddings and classifier logits, enabling alignment of semantic structure in addition to domain-level features; and (ii) entropy minimization and consistency regularization are applied to stabilize adversarial training" (Sec. IV-C), but Table IV compares the full package against vanilla ADDA, which lacks both. Without an ablation, the reported gains could stem largely from entropy minimization and consistency rather than the weather-aware discriminator.
A second concern is the unexplained per-class pattern on the target set. WA-ADDA with ResNet-50 achieves only 63.8% on Clear yet 100.0% on Fog and 97.7% on Snow (Table IV); similar inversions appear for ADDA. The paper does not discuss why the easiest class degrades the most after adaptation. Finally, the target "Street-View" dataset is not attributed to a public source, and the paper omits comparisons to other standard UDA baselines (e.g., DANN, CORAL, CDAN), which limits the evidence that the discriminator conditioning itself is the decisive factor.
The quantitative evidence supports that WA-ADDA outperforms Source-Only and ADDA on the authors' benchmark: "Both ADDA and the proposed WA-ADDA significantly outperform the Source-Only baseline across all backbones, highlighting the necessity of adversarial domain adaptation" (Sec. IV-D). However, because WA-ADDA adds multiple components simultaneously, the comparison to ADDA is confounded. The paper also misses stronger baselines; for instance, DANN and CDAN are standard references here and should have been included. The consistency term uses "mild, differentiable adverse-style perturbations $\mathcal{A}$ (e.g., light fog/rain blur, contrast reduction)" (Sec. III-C), but no implementation details are given, so the evidence for this term is opaque.
The authors provide a GitHub repository with split files and training scripts (Sec. IV-B), and report means over three random seeds. Hyperparameters are listed in Table II. However, the target "Street-View" images are not redistributed: "No original images are redistributed; the repository provides only split files/identifiers and metadata. Raw data must be obtained from the original dataset providers" (Sec. IV-A footnote). Because the target dataset itself is not attributed to a public source, independent reproduction is blocked unless one can reconstruct the exact target collection. Details of the differentiable augmentations $\mathcal{A}$ are also omitted, and learning rate schedules are not specified, leaving gaps for exact replication.
Adverse conditions such as rain, snow, fog, and dust remain challenging for camera-based perception in autonomous driving. We study multi-class weather recognition from street-view images under domain shift, where most available training data come from non-street-view sources that differ markedly from real driving scenes. We propose Weather-Aware Adversarial Discriminative Domain Adaptation (WA-ADDA), which conditions the domain discriminator on predicted weather to promote features that are both domain-invariant and weather-sensitive. We also assemble a multi-dataset benchmark by unifying diverse non-street-view weather collections as sources and real street-view images as targets, and define a standardized evaluation protocol with macro accuracy as the primary metric. Across backbones (ResNet-50, EfficientNet, VGG, DenseNet), WA-ADDA consistently improves street-view performance and yields strong per-class recalls in challenging conditions while preserving clear-weather accuracy. These findings highlight the feasibility of domain-adapted weather recognition and the value of our benchmark for advancing robust, on-board perception.
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.