本文提出了一种弱监督学习方法，用于训练拼接网络创建真实世界的360度全景图像，解决了深度学习拼接模型中获取窄视场和真实图像输入的难题。模型包括颜色一致性校正、扭曲和混合，通过感知和SSIM损失训练，有效处理深度差异和颜色不一致导致的视差失真。实验表明，该方法在真实世界拼接数据集上优于传统方法和现有深度学习模型。
第一次来请先看这篇文章：【图像拼接（Image Stitching）】关于【图像拼接论文精读】专栏的相关说明，包含专栏使用说明、创新思路分享等（不定期更新）
图像拼接系列相关论文精读
- Seam Carving for Content-Aware Image Resizing
- As-Rigid-As-Possible Shape Manipulation
- Adaptive As-Natural-As-Possible Image Stitching
- Shape-Preserving Half-Projective Warps for Image Stitching
- Seam-Driven Image Stitching
- Parallax-tolerant Image Stitching
- Parallax-Tolerant Image Stitching Based on Robust Elastic Warping
- SuperPoint: Self-Supervised Interest Point Detection and Description
- As-Projective-As-Possible Image Stitching with Moving DLT
- Leveraging Line-point Consistence to Preserve Structures for Wide Parallax Image Stitching
- Single-Perspective Warps in Natural Image Stitching
- Local-Adaptive Image Alignment Based on Triangular Facet Approximation
- Image stitching method by multi-feature constrained alignment and colour adjustment
- Natural Image Stitching Using Depth Maps
- Perception-based seam cutting for image stitching
- Stable Linear Structures and Seam Measurements for Parallax Image Stitching
订阅专栏 解锁全文
3万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
