EVENT DOSSIER
The latest research proposes a depth estimation model called ‘Weather-Conditioned Depth Anything’, aimed at addressing the issue of reduced performance of existing methods in variable weather conditions. By incorporating weather conditions as input features, this model significantly improves the accuracy and stability of depth predictions in harsh environments such as rain, snow, and fog. Experiments show that this method effectively overcomes the failure caused by lighting changes or precipitation occlusion in traditional algorithms, providing a more reliable solution for application scenarios that rely on three-dimensional perception, such as autonomous driving, robot navigation, and remote sensing monitoring.
研究人员提出 Weather-Conditioned Depth Anything (DA-W) 框架，旨在解决单目深度估计模型在雾、雨、雪及夜间等恶劣天气下性能失效的问题。该框架通过训练风格过滤器提取内容无关的退化感知天气嵌入，并利用参数高效适配器将其注入 Depth Anything 骨干网络。实验表明，DA-W 在 curated weather benchmarks 上平均提升 AbsRel 指标 3.7%，达到最先进的鲁棒深度估计水平，同时在标准清洁基准测试中表现持平或略优。