Weather-Aware Domain Adaptation for Street-View Weather Recognition

Hossein Maghsoumi; George Atia; Yaser P. Fallah

cs.CVOct 1, 2026

Weather-Aware Domain Adaptation for Street-View Weather Recognition

Authors: Hossein Maghsoumi, George Atia, Yaser P. Fallah

Organizations: Department of Electrical and Computer Engineering University of Central Florida, Orlando, USA · Department of Computer Science University of Central Florida, Orlando, USA

Abstract

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.

Figures & tables

Fig. 1: Overview of the proposed Weather-Aware ADDA (WA-ADDA). Step (1): Pretrain source encoder $M_{s}$ and classifier $C$ on labeled source data. Step (2): Initialize target encoder $M_{t}\leftarrow M_{s}$ and train adversarially with discriminator $D_{\phi}$ . The discriminator conditions on class probabilities $p=C(z)$ in addition to features $z$ , aligning domains within weather modes . Optional entropy minimization and adverse-style consistency further stabilize target predictions. During inference, only $(M_{t},C)$ are retained for weather classification on street-view images.

\mathcal{L}_{M_{t}}=\mathcal{L}_{\mathrm{adv}}^{t}+\lambda_{\mathrm{ent}}\mathcal{L}_{\mathrm{ent}}^{t}+\lambda_{\mathrm{cons}}\mathcal{L}_{\mathrm{cons}}^{t}.

Algorithm 1 Training Weather-Aware ADDA (WA-ADDA)

WA-ADDA Dataset [ 13 ]	Clear	Rain	Snow	Fog	Sand
Source (Non-Street-View)	3916	1310	1765	1400	714
Target (Street-View)	370	379	386	382	369

TABLE I: Distribution of samples across weather categories for both source and target domains.

Backbone	$lr_{mt}$	$lr_{disc}$	$\lambda_{\text{da}}$	$\lambda_{\text{ent}}$	$\lambda_{\text{cons}}$	Disc Mode
ResNet-50	$2\times 10^{-5}$	$5\times 10^{-6}$	$0.010$	$5\times 10^{-4}$	$2\times 10^{-3}$	wa
EfficientNet-B0	$1\times 10^{-5}$	$1\times 10^{-5}$	$0.015$	$2\times 10^{-4}$	$1\times 10^{-3}$	wa
VGG16-BN	$2\times 10^{-5}$	$5\times 10^{-6}$	$0.010$	$5\times 10^{-4}$	$2\times 10^{-3}$	wa
DenseNet-121	$1.5\times 10^{-5}$	$5\times 10^{-6}$	$0.010$	$5\times 10^{-4}$	$2\times 10^{-3}$	wa

TABLE II: Hyperparameter configuration for each backbone during adversarial adaptation.

Method	Backbone	Macro	Overall	Clear	Fog	Rain	Sand	Snow
Source-Only	ResNet-50	94.4	95.5	97.4	95.4	92.2	91.2	95.8
	DenseNet-121	94.5	95.7	97.6	95.4	93.0	90.0	96.4
	EfficientNet-B0	95.1	96.3	97.9	94.1	96.9	88.7	97.6
	VGG16-BN	93.7	95.4	98.2	94.8	89.1	88.7	97.6

TABLE III: Macro accuracy, overall accuracy, and per-class accuracy (%) on the Source dataset.

Method	Backbone	Macro	Overall	Clear	Fog	Rain	Sand	Snow
Source-Only	ResNet-50	68.4	68.5	43.8	90.6	51.4	83.7	72.5
	DenseNet-121	74.3	74.2	78.1	94.2	49.3	85.6	64.0
	EfficientNet-B0	62.4	62.1	64.3	69.6	57.3	92.7	28.0
	VGG16-BN	58.3	58.1	59.2	73.6	56.0	77.6	25.1
ADDA [ 18 ]	ResNet-50	73.0	73.1	57.8	78.8	51.9	85.9	90.7
	DenseNet-121	83.8	83.9	63.6	98.9	76.2	90.0	90.4

TABLE IV: Macro accuracy, overall accuracy, and per-class accuracy (%) on the Street-View target dataset.

Panel of Figure 7 — Fig. 2: Variance across seeds (mean $\pm$ std) on the Street-View target set.

Fig. 3: t-SNE visualizations of source and target feature distributions for different backbones before adaptation (source-only) and after adaptation using ADDA and the proposed WA-ADDA. Each column corresponds to one backbone, and each row corresponds to one training setting.

Explore similar work

WILD SAM: A Simulated-and-Real Data Augmentation for Autonomous Driving Perception under Challenging Weather
May 1, 2026Hamed Khatounabadi, Xiaohu Lu, Hayder RadhaAutonomous Driving Perception Domain Adaptation
Electrical and Computer Engineering Michigan State University East Lansing, USA
XWOD: A Real-World Benchmark for Object Detection under Extreme Weather Conditions
May 12, 2026Chih-Hsin Chen, Yu-Tung Liu, Amar Fadillah +2Object Detection Cross-Dataset Benchmark
Department of Electronic Engineering, National Taipei University of Technology · Adobe Inc.
ObsDriveBench: Benchmarking Multimodal Understanding under Adverse Weather with Observability Awareness
Jul 26, 2026Qiao Yan, Yihan Wang, Zhenghao Xing +2Autonomous Driving Multimodal Benchmarks
Department of Computer Science and Engineering, The Chinese University of Hong Kong · Institute of Medical Intelligence and XR, The Chinese University of Hong Kong

May 1, 2026, cs.CVJ/K move · Enter open · S save

WILD SAM: A Simulated-and-Real Data Augmentation for Autonomous Driving Perception under Challenging Weather

Hamed Khatounabadi, Xiaohu Lu, Hayder Radha

Electrical and Computer Engineering Michigan State University East Lansing, USA

Autonomous Driving Perception Domain Adaptation Data Augmentation Mitigating Domain Shift Iterative Denoising Methods

Open paper AskPDF arXiv

The performance of state-of-the-art object detectors degrades significantly under adverse weather, causing a safety-critical domain shift problem for autonomous vehicles. Recent efforts address this problem by relying on synthetic data to train the object detectors, which limits their real-world applicability. Meanwhile, pseudo-labeling is widely used for cross-dataset domain adaptation problems. However, these methods have not been exploited by weather-based domain adaptation approaches due to the noisy nature of such labels generated under harsh weather conditions. In this paper, we propose two new approaches to mitigate this weather-induced domain shift. First, we propose a Weather-Induced pseudo Label Denoising (WILD) framework that filters noisy pseudo labels generated by real data captured under adverse weather conditions. Second, we develop a novel hybrid training methodology, WILD SAM, that exploits both pseudo-label denoising and simulation-based training solutions while using real-data from the target harsh-weather domain. We validate both proposed approaches, WILD and WILD SAM, on the recently released Four Seasons dataset across rainy and snowy scenarios. Experiments show that the proposed frameworks improve Average Precision (AP) up to 13% and significantly reduce the weather-induced performance gap relative to the baseline. The code is available at: https://github.com/Kh-Hamed/WILD-SAM

Abstract

Explore similar work

WILD SAM: A Simulated-and-Real Data Augmentation for Autonomous Driving Perception under Challenging Weather

XWOD: A Real-World Benchmark for Object Detection under Extreme Weather Conditions

ObsDriveBench: Benchmarking Multimodal Understanding under Adverse Weather with Observability Awareness