Video Desnower: An Adaptive Feature Fusion Understanding Video Desnowing Model With Deformable Convolution and KNN Point Cloud Transformer

Abstract

The desnowing (snow removal) model is extensively utilized in various fields, including visual enhancement, security monitoring, and autonomous driving technology. Some previous work developed highly efficient models that primarily addressed single-image desnowing tasks. Simultaneously, the process of video desnowing holds significance in practical applications. There is only a limited amount of literature available on the topic of video desnowing, mainly utilizing predetermined knowledge rather than exploring deep learning technologies. Given the identified deficiency in current research, our study aims to improve upon existing video desnowing methodologies by introducing an innovative approach and filling the void of specialized datasets. Our contribution includes the development of a dataset tailored for the training and assessment of video desnowing models, as well as the creation of the Video-Denower model, which integrates adaptive feature fusion mechanisms. Video-Desnower employs sophisticated adaptive feature fusion methodologies to enhance desnowing efficacy through the comprehensive analysis of features across various scales. In contrast to single-image models, this particular model has the ability to analyze multiple frames within a video. Experiments on a video desnowing dataset show its exceptional capabilities. The code and dataset used in this study are available upon request. Interested researchers can contact us at for access. Please include a brief description of your research interest and how you intend to use the data.

Keywords

• Computer vision
• deep learning
• video desnowing
• feature fusion understanding