Supervised Small-Baseline and Large-Baseline Homography Learning With Diffusion-Based Data Generation

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Homography estimation is crucial for tasks like image stitching and augmented reality.
• Supervised learning methods for homography estimation require large, accurately labeled datasets, which are difficult to obtain.

Purpose of the Study:

• To propose an iterative framework for generating realistic training data for supervised homography learning.
• To develop a state-of-the-art homography estimation network using the generated data.

Main Methods:

• An iterative framework with distinct generation and training phases.
• Data generation involves using pre-estimated masks and homographies, along with sampled ground truth homographies.
• Training phase refines data using a content refinement diffusion model and iteratively updates the homography network.

Main Results:

• The proposed method achieves state-of-the-art performance in homography estimation.
• The iterative strategy simultaneously improves dataset quality and network performance.
• Existing supervised homography methods benefit from the generated dataset.

Conclusions:

• The iterative framework effectively generates high-quality training data for homography learning.
• This approach leads to superior homography estimation network performance.
• The method offers a viable solution for creating realistic datasets in computer vision.