Research map / 3D vision / Depth and stereo
Depth-Supervised Fusion Network for Seamless-Free Image Stitching
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a depth-supervised image stitching network (DSFN) that uses Depth Anything priors for two-stage depth-aware alignment, soft-seam diffusion fusion, and reparameterized shift regression to handle large parallax.
Problem
Image stitching under large parallax from significant depth variation produces ghosting and misalignment; existing homography and deep methods struggle with multi-depth scenes and cross-domain generalization.
Approach
Injects relative depth from Depth Anything as a supervision prior for transformation estimation (coarse homography plus RBF mesh refinement), replaces seam cutting with a learned soft-seam mask diffused from graph-based seam costs, and uses a reparameterized RepBlock with adaptive pruning of the 1x1 branch in the regression head.
Key results
- On UDIS-D, the method reports PSNR 25.467, SSIM 0.839, SIQE 43.732, LPIPS 0.462, edging out UDIS++ (25.426, 0.837).
- On IVSD, PSNR 26.778 and SSIM 0.820 versus UDIS++ 26.649 and 0.819.
- Runs at 67.04 ms per 512x512 image versus 79.73 ms for UDIS++ (fastest reported).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed depth-supervised fusion method instead of UDIS++ | Image stitching · UDIS-D | PSNR: 25.426 -> 25.467 | Helped |
| Use the proposed depth-supervised fusion method instead of UDIS++ | Image stitching · UDIS-D | SSIM: 0.837 -> 0.839 | Helped |
| Use the proposed depth-supervised fusion method instead of LPC | Image stitching · UDIS-D | SIQE: 43.616 -> 43.732 | Helped |
| Use the proposed depth-supervised fusion method instead of SPW | Image stitching · UDIS-D | LPIPS: 0.466 -> 0.462 | Helped |
| Use the proposed depth-supervised fusion method instead of UDIS++ | Image stitching · IVSD | PSNR: 26.649 -> 26.778 | Helped |
| Use the proposed depth-supervised fusion method instead of UDIS++ | Image stitching · IVSD | SSIM: 0.819 -> 0.820 | Helped |
| Use the proposed depth-supervised fusion method instead of UDIS++ | Image stitching · IVSD | SIQE: 46.383 -> 46.568 | Helped |
| Use the proposed depth-supervised fusion method instead of ELA | Image stitching · IVSD | LPIPS: 0.435 -> 0.436 | Hurt |
| Use the proposed method instead of UDIS++ | Image stitching · 512×512 image stitching | Time (ms): 79.73 -> 67.04 | Helped |
| Add mesh constraints (ablation) | Depth-aware transformation estimation · UDIS-D | Visual distortion | Helped |
| Add depth supervision (ablation) | Depth-aware transformation estimation and soft-seam fusion · UDIS-D | Overall accuracy | Helped |
| Add fusion-module constraints (ablation) | Soft-seam multi-view fusion · UDIS-D | Compensation for stitching errors | Helped |
Surprising findings
- Full reparameterization and the original model both underperform an intermediate threshold (c-hat = 0.25) in the RBA ablation.
Limitations
- Designed for two-image stitching; lacks robust multi-image panoramic stitching and loop consistency.
- Depth estimation is unreliable in scenes with dynamic objects.
- Improvements over UDIS++ on UDIS-D are small in most metrics.
Experimental setup
Datasets: UDIS-D
Benchmarks: UDIS-D IVSD
Metrics: PSNR SSIM SIQE LPIPS processing time
Baselines: APAP ELA LPC SPW UDIS UDIS++ TRIS SRS
Builds on: Depth Anything UNet UDIS++ RepVGG
Closest papers by meaning
- Pixel-Wise Warping for Deep Image Stitching (AAAI 2023 · similarity 0.89)
- Reconstructing the Image Stitching Pipeline: Integrating Fusion and Rectangling into a Unified Inpainting Model (NeurIPS 2024 · similarity 0.87)
- Object-level Geometric Structure Preserving for Natural Image Stitching (AAAI 2025 · similarity 0.86)
- Towards Robust Image Stitching: An Adaptive Resistance Learning against Compatible Attacks (AAAI 2024 · similarity 0.83)
- MultiDiff: Consistent Novel View Synthesis from a Single Image (CVPR 2024 · similarity 0.83)
- EditSplat: Multi-View Fusion and Attention-Guided Optimization for View-Consistent 3D Scene Editing with 3D Gaussian Splatting (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.