Core Concepts
A novel image stitching method that combines the fusion and rectangling stages into a single inpainting-based process, leveraging a pre-trained large-scale generative model and a weighted mask to enhance fault tolerance and simplify the stitching pipeline.
Abstract
The paper presents a novel approach to the image stitching problem, which typically involves three distinct stages: registration, fusion, and rectangling. The authors identify limitations in the current sequential pipeline, where errors can propagate through the stages and complicate parameter tuning.
To address these issues, the authors propose SRStitcher, a method that integrates the fusion and rectangling stages into a unified inpainting-based process. The key insights are:
Fusion and rectangling can be effectively combined as a single inpainting problem, where fusion requires adjusting the original image information in the overlapping areas, while rectangling focuses on filling in missing areas.
By employing a pre-trained large-scale generative model and a weighted mask, SRStitcher can solve the fusion and rectangling problems in a single inference, without the need for additional training or fine-tuning.
This approach simplifies the stitching pipeline and enhances fault tolerance towards misregistration errors, as the powerful generalization capabilities of the pre-trained model can produce visually appealing results even with less precise registration.
The authors conduct extensive experiments on the UDIS-D dataset, demonstrating that SRStitcher outperforms state-of-the-art methods in both quantitative assessments and qualitative evaluations. The method also exhibits enhanced robustness in handling challenging scenarios, such as scenes with registration errors, repetitive textures, and large missing regions.
Stats
The stitched image is obtained by overlaying the less distorted image on top of the more distorted one, guided by the UDIS++ registration method.
The weighted mask is designed to control the degree of inpainting required for different regions of the image, with higher intensity in the seam areas and lower intensity in the non-overlapping regions.
Quotes
"By employing the weighted mask and large-scale generative model, SRStitcher can solve the fusion and rectangling problems in a single inference, without additional training or fine-tuning of other models."
"Our method not only simplifies the stitching pipeline but also enhances fault tolerance towards misregistration errors."