A Robust Semantic Segmentation Pipeline for the CVPR 2026 8th UG2+ Challenge Track 2

Abstract: This report presents our solution for the WeatherProof Dataset Challenge, namely CVPR 2026 8th UG2+ Challenge Track 2: Semantic Segmentation in Adverse Weather. For the semantic segmentation task under adverse weather conditions, we propose a semi-supervised segmentation pipeline. Our method is trained exclusively on the WeatherProof dataset, without using any additional external data. Specifically, we adopt UniMatch V2 as the baseline model and treat all degraded-weather images as unlabeled data for semi-supervised training, thereby fully exploiting the data distribution provided by the challenge. During inference, we further apply test-time augmentation to improve the robustness and segmentation accuracy of the final predictions. The code is publicly available at: https://github.com/ylb888/weatherproof-challenge-unimatchv2.

• The paper demonstrates a significant mIoU boost (+0.10) by leveraging weather-degraded data in a semi-supervised setup.
• It employs a teacher-student paradigm with a frozen DINOv2 encoder and strong augmentation techniques to ensure robust feature learning.
• Test-time augmentation is used to further stabilize predictions, confirming improved performance in challenging weather conditions.

Robust Semi-Supervised Semantic Segmentation for Adverse Weather: A Technical Summary

Problem Context and Dataset

Adverse weather significantly degrades the performance of semantic segmentation models, especially foundation models that excel under standard visual conditions. The WeatherProof dataset, with its paired clear-weather and weather-degraded images accompanied by reliable semantic labels, enables systematic evaluation and training for robust segmentation in such challenging conditions. This work addresses semantic segmentation for the CVPR 2026 8th UG2+ Challenge Track 2, utilizing only the WeatherProof dataset to avoid external domain biases and maximize weather robustness.

Methodology

Semi-Supervised Teacher-Student Architecture

The approach is based on a semi-supervised teacher-student paradigm, leveraging both supervised training on clean images and consistency regularization on degraded (unlabeled) images. The architecture consists of:

• Backbone Encoder: DINOv2-Base, frozen during training to retain foundation model generality.
• Segmentation Framework: DPT architecture with a task-specific segmentation head.
• Student Network: Updated via gradient descent.
• Teacher Network: An EMA (exponential moving average) copy of the student, used for generating stable pseudo labels.

Training Objectives

• Supervised Loss: Cross-entropy on clean-weather images with ground-truth labels.
• Unsupervised Consistency Loss: Pseudo-labels for degraded images are generated using weak augmentation passed through the teacher network. Only high-confidence pixels ($T=0.95$ threshold) are used for loss computation.
• Weak-to-Strong Consistency: Strong augmentations (color jitter, blur, CutMix, aggressive cropping) are applied to degraded images before feeding them to the student. Complementary dropout (channel-wise masks with scaling) is used to enforce robust feature learning.
• Overall Loss: Combined supervised and unsupervised losses, with a scalar $\lambda$ weighting the consistency term.

Test-Time Augmentation (TTA)

During inference, standard TTA techniques are applied to boost robustness and accuracy further, aggregating predictions over spatial and photometric transforms to mitigate overfitting to specific weather artifacts.

Experimental Results

Training and Evaluation Protocol

• Training exclusively on the WeatherProof dataset, avoiding any additional external data.
• Images cropped to $518 \times 518$ pixels.
• DINOv2 encoder remained frozen to ensure backbone representations remain generic.
• 60 epochs of training with a batch size of 32 on dual NVIDIA V100 GPUs.

Quantitative Results

The ablation experiments demonstrate the effectiveness of the methodology:

Training Setting	mIoU	mDice
Clean only	0.69	0.69
Clean + Degraded	0.79	0.79
Clean + Degraded + Test-Time Aug	0.80	0.80

The most significant performance gain ($+0.10$ mIoU/mDice) results from incorporating degraded images via the semi-supervised framework. TTA confers an additional, but smaller, improvement.

Claims

• The model achieves a robust mIoU of 0.80 on adverse weather segmentation, trained solely on the provided dataset.
• Semi-supervised consistency learning using weather-degraded images yields a substantial mIoU improvement over supervised-only training.
• Test-time augmentation provides marginal gains, supporting its use for stability under domain shifts.

Implications and Future Directions

This work affirms that, even absent external data, semi-supervised architectures leveraging both labeled and unlabeled degradations significantly enhance segmentation robustness under challenging conditions. By adopting complementary dropout and carefully aligned augmentations, the approach improves generalization to diverse weather scenarios while maintaining efficient training by freezing powerful foundation encoders.

Practically, this technique is directly relevant to domains like autonomous driving and robotics, where weather resilience is a critical requirement. Theoretically, the results emphasize the utility of teacher-student consistency frameworks, particularly when annotation scarcity coincides with an abundance of challenging unlabeled data.

Possible future avenues include:

• Extending the framework to exploit temporal cues from video data in adverse conditions.
• Incorporation of generative data augmentation to synthesize additional degradations.
• Adaptive confidence thresholding for pseudo-labels based on domain-specific uncertainty estimations.
• Investigating end-to-end backbone adaptation strategies without compromising base representation quality.

Conclusion

This paper presents a robust semi-supervised pipeline for semantic segmentation in adverse weather, demonstrating substantial improvements in the WeatherProof benchmark solely through architectural and training innovations. The framework effectively exploits paired clean and degraded scenes, establishing a strong baseline for research targeting weather-robust perception without domain leakage or external auxiliary datasets.