Reconstructing the Image Stitching Pipeline: Integrating Fusion and Rectangling into a Unified Inpainting Model

Deep learning-based image stitching pipelines are typically divided into three cascading stages: registration, fusion, and rectangling. Each stage requires its own network training and is tightly coupled to the others, leading to error propagation and posing significant challenges to parameter tuning and system stability. This paper proposes the Simple and Robust Stitcher (SRStitcher), which revolutionizes the image stitching pipeline by simplifying the fusion and rectangling stages into a unified inpainting model, requiring no model training or fine-tuning. We reformulate the problem definitions of the fusion and rectangling stages and demonstrate that they can be effectively integrated into an inpainting task. Furthermore, we design the weighted masks to guide the reverse process in a pre-trained largescale diffusion model, implementing this integrated inpainting task in a single inference. Through extensive experimentation, we verify the interpretability and generalization capabilities of this unified model, demonstrating that SRStitcher outperforms state-of-the-art methods in both performance and stability. Code: https://github.com/yayoyo66/SRStitcher

Paper

References (56)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer yn5D5/10 · confidence 5/52024-07-08

Summary

This work proposes to integrate the fusion and rectangling of image stitching into a unified inpainting model. In particular, the weighted masks are designed to guide the reverse process in a pre-trained large-scale diffusion model, which implements this integrated inpainting task in a single inference. Extensive experiments demonstrate the interpretability and generalization capabilities of the proposed unified model.

Strengths

+ The motivation for reconstructing the image stitching pipeline into a unified model is sound and clear. + Some discussions on image stitching are insightful. For example, in the Introduction, the authors claimed that "To address the error propagation problem, we identify image fusion as the key point for improvement". I agree it is an accurate point to inspire this work. + The reviews of previous methods are detailed and comprehensive (in Tab. 1 and Section B in the Appendix), which clearly shows their limitations and potential improvements.

Weaknesses

- While this work is well-motivated, the proposed contributions seem to be weak for me. For instance, the authors proposed a weighted mask to guide the reverse process in the diffusion model. However, the current presentation regarding this contribution looks like an experimental trick rather than a technical proposal. - The contribution of the proposed unified inpainting model (Section 3.1) is also expected to be highlighted. It is still ambiguous how the fusion part and rectangling part are merged into a unified model. I believe this module is much more important than applying cutting-edge generative models. - This work leverages a **generation** view to address the **reconstruction** problem. Such an interesting trend can be noticed in the recent works from other research regions as well. However, how to balance the generation and reconstruction in the image stitching problem remains unclear. I believe the users may prefer accurate reconstruction instead of introducing novel content (sometimes artifacts). Moreover, some previous works also proposed to fill the irregular boundaries of images or holes by warping with generative models, such as "Free View Synthesis (ECCV'2020)", "Towards complete scene and regular shape for distortion rectification by curve-aware extrapolation (ICCV'2021)", "iNVS: Repurposing Diffusion Inpainters for Novel View Synthesis (SIGGRAPH Asia)", etc. More discussions are expected to be provided. - There are some typos in the paper: Line 49: "Therefore, We question". The authors are suggested to further polish this work.

Questions

- Can the proposed unified model be extended into other research areas?

Rating

5

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

- The proposed unified inpainting model seems to be impractical compared to previous reconstruction-based solutions.

Reviewer pLdt6/10 · confidence 4/52024-07-10

Summary

This paper proposes an image stitching algorithm that unifies fusion and rectanguling stages of a conventional pipeline with an image inpainting diffusion model applied with a progressive reverse process guided by weighted masks. Their reverse algorithm progressively inpaints seam regions by gradually increasing the size of soft mask holes, while keep outpainting rectanguling regions with the same hard mask holes. The proposed work is compared with previous works by keeping the registration and fusion stages with UDIS/UDIS++ while altering with various models for a rectangling stage.

Strengths

In their setup of experiments, theirs generally showed higher quality quantitatively and quantitatively than compared methods. The proposed reverse process results in a high quality generation, suitable to inpaint and outpaint the seam and rectangling regions naturally.

Weaknesses

It is unclear if the repeated reverse process with gradually dilating mask holes is really beneficial than a single reverse process guided by a fixed mask. The repeated application of the re-noising + de-noising step makes the algorithm much slower. Considering the seam regions are usually narrow while the rectanguling regions could be quite wide, the progressive soft mask dilation during the progressive reverse process may be more useful with the rectanguling masks than the seam masks.

Questions

Comparing with a case of a single reverse process guided by a fixed mask would make the paper more convincing. Comparing with a case that applies progressive mask dilation of the rectanguling mask holes would be useful.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Technical limitations are properly addressed. No serious potential negative societal impact is expected.

Reviewer pLdt2024-08-12

Thank you for your rebuttal. It clarifies the need of different masking strateges for the seam and rectangling regions, and the preservation of computational complexity for the inference. Therefore, my major concerns have been resolved. I find that the proposed scheme, which converts conventional multiple stages for a stiching problem into unified one by adapting diffusion based inpainting models, holds certain values suitable for the venue. Existing inpatining models would not be able to achieve this goal without applying the proposed algorithm. I would like to update the final rating from from BA to WA.

Reviewer m57w6/10 · confidence 4/52024-07-13

Summary

The paper introduces SRStitcher, a novel method that integrates the fusion and rectangling stages of the image stitching pipeline into a unified inpainting model using a pre-trained large-scale diffusion model, eliminating the need for additional training. This approach addresses the issue of error propagation in traditional pipelines, offering a streamlined and robust solution. Strengths of the method include improved performance in image quality and content consistency, as well as robustness to registration errors. The experimental results are extensive, providing both quantitative and qualitative evidence of SRStitcher's superiority over existing state-of-the-art methods.

Strengths

SRStitcher's primary strength lies in its innovative integration of the fusion and rectangling stages into a unified inpainting model, which addresses the long-standing issue of error propagation in traditional image stitching pipelines. By leveraging a pre-trained large-scale diffusion model, SRStitcher eliminates the need for stage-specific training, thereby simplifying the pipeline and enhancing its robustness. This approach ensures superior performance in handling registration errors, which are typically propagated and amplified in multi-stage pipelines. Additionally, the use of weighted masks to guide the inpainting process allows for precise control over inpainting intensity, significantly improving image quality and content consistency. These advancements address the limitations of existing methods, which often struggle with the independent optimization of each stage and the associated parameter tuning challenges. The extensive experimental results, including both quantitative metrics and qualitative assessments, robustly demonstrate SRStitcher's superiority over state-of-the-art methods, showcasing its ability to produce high-quality stitched images with greater stability and fewer artifacts.

Weaknesses

While SRStitcher simplifies the image stitching pipeline by integrating the fusion and rectangling stages into a unified inpainting model, its technical novelty is questionable. The approach primarily combines existing technologies—pre-trained diffusion models and weighted masks—rather than introducing fundamentally new methodologies. The integration, while effective, does not inherently surpass the capabilities of current state-of-the-art techniques used separately for fusion and rectangling. For SRStitcher to be considered truly novel, it should achieve technical goals that were unattainable with the two processes handled independently. As it stands, the method appears to be more of a consolidation of existing practices rather than a groundbreaking innovation, merely reorganizing the workflow without providing substantial new capabilities or overcoming significant limitations of the previous approaches.

Questions

- Technical Novelty: How does SRStitcher fundamentally advance the field of image stitching beyond merely integrating existing fusion and rectangling processes into a unified model? Can the authors provide more evidence of novel technical contributions? - Performance Comparison: While the paper claims superior performance, how does SRStitcher specifically outperform existing state-of-the-art methods in scenarios with extreme registration errors(large parallax)? Are there any edge cases where SRStitcher struggles compared to traditional methods?

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have adequately addressed the limitations of their work and there appear to be no significant issues with the broader societal impacts. They have demonstrated transparency and responsibility in discussing the constraints and potential improvements of SRStitcher. Overall, their approach seems robust and well-considered, with no apparent concerns regarding its application or societal implications.

Reviewer GE1w6/10 · confidence 4/52024-07-14

Summary

This paper tried to integrate the fusion and rectangling stages in image stitching into a unified model. More concretely, a special fusion, a rectanlging step, and a mask-guided diffusion model are gathered to implement stitching-customized image inpainting, especially for the irregular boundaries. It is worth mentioning that, the inpainting model uses the pre-trained model and requires no more fine-tuning. To evaluate the proposed method, the authors designed a quantitative metric named CCS and conducted extensive experiments.

Strengths

1. The concept of integrating multiple stages of image stitching into a single stage is novel and promising. 2. The experiments are abundant and convincing. 3. The authors leverage a pre-trained model to implement stitching-customized inpainting without any extra training.

Weaknesses

1. The so-called unified model consists of several steps including a specially-designed fusion step, a rectangling step, and a inpainting step. I don’t think this multi-step design can be regarded as a unified model. 2. I am skeptical of whether the inpainting model is meaningful to image stitching. As claimed in the DeepRectangling paper, they abandoned the inpainting model because they think it may introduce some contents that are far from the reality. The manuscript seems to ignore the problem. And I do not think the proposed method can address this issue in the so-called unified inpainting model. 3. The inpainting results are still not perfect, as illustrated in the image boundaries of Fig. 14. 4. The fusion step of Eq. 4 is special. What’s the motivation for that?

Questions

My concern lies in my second weakness. Will this inpainting model for irregular boundaries be meaningful? I have another idea for the inpainting model of image stitching. The inpainting model should not concentrate on the boundaries but on the regions where the artifacts and distortion are produced. Let’s put it in this way. The registration stage may introduce artifacts or distortion. So can we eliminate these issues through an inpainting model? It may be more meaningful to locate these regions, mask them, and then inpaint them. The inpainting process can be implemented with the guidance of original contents, thus contributing to a reliable completion model. This is just a simple discussion. But from my perspective, I cannot figure out the meaning of the proposed inpainting model.

Rating

6

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

No limitation is mentioned in the manuscript.

Authorsrebuttal2024-08-13

Dear Area Chair, We write this comment in the hope that you can remind the Reviewers to reply to the Rebuttal. The time is running out, but only one Reviewer has replied now (Thanks for the timely reply from Reviewer pLdt). According to Notifications, we as authors cannot rush Reviewers to reply, which should be the responsibility of the Area Chair. We also do not know whether the Area Chair has reminded Reviewers of the deadline, because we have not received any Notification from Area Chair. Since there is only one day left and there are still three reviewers who have not replied, we hope the Area Chair will remind the reviewers of the deadline. Whether the score is changed or not, the reviewers' reply to our Rebuttal will be a good reference for our subsequent work. We believe that this is why Neurips started the discussion session. We don't think a discussion session where most people are silent is what Neurips hopes to see.

Area Chair h8Lo2024-08-13

reminder was sent

Dear authors, The reminders were sent to individual reviewers without posting to the public. I hope that they will post soon. AC

Reviewer yn5D2024-08-14

Thanks for the detailed rebuttal, which addressed most of my concerns. I would like to increase my rating. The updated clarifications and experiments are expected to be presented in the final version.

Reviewer GE1w2024-08-14

My concerns are addressed in the rebuttal and I would like to raise my evaluation.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC