Depth-Supervised Fusion Network for Seamless-Free Image Stitching

TL;DR

This work tackles ghosting and misalignment in image stitching caused by large depth-induced parallax by introducing a depth-consistency constrained framework. It combines a depth-aware two-stage transformation estimation with a diffusion-based soft-seam fusion and a reparameterization regression block to improve both alignment accuracy and computational efficiency. Depth priors from monocular depth models regularize the alignment, while soft-seam diffusion yields natural transitions and reduces artifacts. Across UDIS-D and IVSD datasets, the method demonstrates superior alignment quality, seamless fusion, and competitive efficiency compared to state-of-the-art approaches.

Abstract

Image stitching synthesizes images captured from multiple perspectives into a single image with a broader field of view. The significant variations in object depth often lead to large parallax, resulting in ghosting and misalignment in the stitched results. To address this, we propose a depth-consistency-constrained seamless-free image stitching method. First, to tackle the multi-view alignment difficulties caused by parallax, a multi-stage mechanism combined with global depth regularization constraints is developed to enhance the alignment accuracy of the same apparent target across different depth ranges. Second, during the multi-view image fusion process, an optimal stitching seam is determined through graph-based low-cost computation, and a soft-seam region is diffused to precisely locate transition areas, thereby effectively mitigating alignment errors induced by parallax and achieving natural and seamless stitching results. Furthermore, considering the computational overhead in the shift regression process, a reparameterization strategy is incorporated to optimize the structural design, significantly improving algorithm efficiency while maintaining optimal performance. Extensive experiments demonstrate the superior performance of the proposed method against the existing methods. Code is available at https://github.com/DLUT-YRH/DSFN.

Figures (9)

• Figure 1: Workflow of the proposed method. It consists of two procedure: depth-aware transformation estimation and soft-seam based multi-view fusion. Besides, the transformation estimation process incorporates reparameterized regression to establish the optimal model.
• Figure 2: Visual comparison of stitched images from UDIS-D dataset. The alignment error is visualized in the lower right corner.
• Figure 3: Visual comparison of stitched images from IVSD dataset. The alignment error is visualized in the lower right corner.
• Figure 4: Visual quality survey.
• Figure 5: Ablation study on different loss components.