MOWA: Multiple-in-One Image Warping Model.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Existing image warping methods require task-specific models, limiting generalization to different camera models and manipulations.
• Current approaches struggle with multi-task learning and adapting to diverse real-world warping scenarios.

Purpose of the Study:

• To develop a unified image warping model capable of handling multiple tasks simultaneously.
• To improve generalization and adaptability in image warping across various camera models and custom manipulations.

Main Methods:

• Proposed a Multiple-in-One image WArping (MOWA) model to address multi-task learning challenges.
• Disentangled motion estimation at both region and pixel levels for robust performance.
• Introduced a lightweight point-based classifier for dynamic, task-aware feature map modulation.

Main Results:

• MOWA, trained on six warping tasks, outperforms state-of-the-art task-specific models on most benchmarks.
• Demonstrated significant generalization potential on unseen scenes through cross-domain and zero-shot evaluations.
• Achieved accurate estimation by leveraging task-type prompts for feature map modulation.

Conclusions:

• MOWA is the first model to successfully address multiple practical image warping tasks within a single framework.
• The proposed approach offers superior performance and generalization compared to existing task-specific image warping solutions.
• MOWA paves the way for more versatile and adaptable image manipulation technologies.