AutoAWG: Adverse Weather Generation with Adaptive Multi-Controls for Automotive Videos
Summary: arXiv:2604.18993v1 Announce Type: cross
Perception robustness under adverse weather conditions remains a critical challenge for autonomous driving systems. One of the primary obstacles is the scarcity of real-world video data that captures such adverse weather scenarios. Traditional weather generation approaches often struggle to achieve a balance between visual quality and the reusability of annotations. Addressing these challenges, we introduce AutoAWG, a controllable Adverse Weather Video Generation framework specifically designed for autonomous driving applications.
Key Features of AutoAWG
- Semantics-guided Adaptive Fusion: AutoAWG employs a unique semantics-guided adaptive fusion of multiple controls. This allows the framework to achieve strong weather stylization while preserving the high-fidelity representation of safety-critical targets in the video.
- Vanishing Point-anchored Temporal Synthesis: The framework leverages a vanishing point-anchored temporal synthesis strategy to construct training sequences from static images. This innovative approach significantly reduces reliance on synthetic data, which has been a limitation in previous methods.
- Masked Training for Stability: AutoAWG adopts a masked training technique that enhances the stability of long-horizon generation, ensuring that the generated video sequences maintain coherence over extended periods.
Performance Metrics
On the nuScenes validation set, AutoAWG has demonstrated substantial improvements over prior state-of-the-art methods. The performance metrics indicate:
- Without first-frame conditioning, the Fréchet Inception Distance (FID) and Fréchet Video Distance (FVD) are reduced by 50.0% and 16.1%, respectively.
- With first-frame conditioning, the FID and FVD show further reductions of 8.7% and 7.2%.
Qualitative and Quantitative Results
Extensive qualitative and quantitative evaluations reveal notable advantages of AutoAWG in several key areas:
- Style Fidelity: AutoAWG maintains high fidelity in the visual style of the generated videos, ensuring that they closely resemble real-world adverse weather conditions.
- Temporal Consistency: The generated sequences exhibit strong temporal consistency, making them suitable for training autonomous driving systems that require stable input over time.
- Semantic-Structural Integrity: The framework ensures that the structural integrity of the scene is preserved, which is vital for safety-critical applications in autonomous vehicles.
Conclusion
The practical value of AutoAWG is evident in its ability to improve downstream perception tasks in autonomous driving by effectively simulating adverse weather conditions. This advancement not only contributes to the robustness of perception systems but also opens up new avenues for research and development in the field of autonomous driving technology. For those interested in exploring AutoAWG further, the source code is available at GitHub.