Core Concepts
A zero-shot omnidirectional image super-resolution method, OmniSSR, leverages the image prior of Stable Diffusion model and employs Octadecaplex Tangent Information Interaction and Gradient Decomposition to achieve high-fidelity and high-quality super-resolution results without any training or fine-tuning.
Abstract
The paper proposes OmniSSR, a zero-shot omnidirectional image super-resolution method that leverages the image prior of the Stable Diffusion (SD) model. The key highlights are:
Octadecaplex Tangent Information Interaction (OTII): The method transforms the input equirectangular projection (ERP) omnidirectional images into tangent projection (TP) images, whose distribution approximates that of planar images. This enables the use of the original SD super-resolution method for planar images. The TP images are then transformed back to the ERP format.
Gradient Decomposition (GD) Correction: To enhance the consistency of the SR results from SD, the method employs a convex optimization-based GD correction technique. This iteratively refines the initial super-resolution results, improving both the fidelity and realness of the restored images.
Zero-shot Approach: The proposed OmniSSR method is training-free, requiring no fine-tuning or specialized training on omnidirectional image datasets. This mitigates the data demand and overfitting issues associated with end-to-end training.
Experiments on benchmark datasets demonstrate the superior performance of OmniSSR compared to existing state-of-the-art omnidirectional image super-resolution methods, in terms of both quantitative metrics and visual quality.
Stats
The degradation for low-resolution ERP images is bicubic down-sampling.
The implementation of the pseudo-inverse of the bicubic downsampling operator is referred from the code of DDRM [25].