Comparison with Inpainting Methods

Image rectangling aims to eliminate irregular boundaries while maintaining as much data consistency and achieving good qualitative results as possible, therefore, previous rectangling methods[1,2] choose to warp stitching images. While inpainting methods are powerful at generating visually pleasing outcomes, they tend to introduce extra content into the stitched ones, as demonstrated below, thus affecting the data consistency negatively. As shown in the table below, inpainted stitching images (row 2 and 3) result in the much lower PSNR/SSIM metrics than by RecDiffusion. Moreover, the quality of the images generated through inpainting techniques does not meet the requirements of this task, as evident from their FID scores. These scores are higher than the FID scores comparing the stitched input images to the ground truth rectangling images, indicating a significant discrepancy in image quality.

Method FID SSIM PSNR
Reference 12.25 0.3245 11.30
Palette[3] - 0.3315 14.49
Stable Diffusion 2.1[4] 15.58 0.3276 14.23
Ours 3.63 0.7733 22.21
References:
[1] Kaiming He, Huiwen Chang, and Jian Sun. Rectangling panoramic images via warping. ACM Trans. Graphics, 32 (4):1–10, 2013.
[2] Laang Nie, Chunyu Lin, Kang Liao, Shuaicheng Liu, and Yao Zhao. Deep rectangling for image stitching: a learning base- line. In Proc. CVPR, pages 5740–5748, 2022.
[3] Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In Proc. ACM SIGGRAPH, pages 1–10, 2022.
[4] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ̈orn Ommer. High-resolution image syn- thesis with latent diffusion models. In Proc. CVPR, pages 10684–10695, 2022.