Research map / Detection and segmentation / Open-set and few-shot detection
Object-level Geometric Structure Preserving for Natural 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 OBJ-GSP, a grid-based image stitching method that uses SAM-segmented object contours with triangle meshes to preserve object shapes while aligning images, and introduces the StitchBench benchmark.
Problem
Image stitching methods achieve good pixel alignment but distort object structures; prior line-based structure preservation fails when lines are absent or misleading.
Approach
Segment objects with SAM-family models, build a triangle mesh from each object's center to sampled contour points, and add an object-level similarity energy term to the GSP grid-based stitching objective so objects keep their shape while alignment is preserved.
Key results
- OBJ-GSP improves mean MDR over GES-GSP by 3.5% and mean NIQE by 3.8% across the benchmark subsets (per the Table 1 caption).
- Against GES-GSP on the GES-GSP dataset, OBJ-GSP reduces MDR from 1.15462 to 0.98288 (improvement reported as 5.9%) and on SPHP dataset NIQE improves 16.0%.
- With SAM ViT-B, ViT-L, ViT-H backbones, MDR improvements over GES-GSP are 1.9%, 2.7%, and 3.6% respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add object-level contour meshes using OBJ-GSP instead of GES-GSP | Image stitching · StitchBench | MDR: 3.5% | Helped |
| Add object-level contour meshes using OBJ-GSP instead of GES-GSP | Image stitching · StitchBench | NIQE: 3.8% | Helped |
| Use OBJ-GSP instead of GES-GSP | Image stitching · SVA | MDR: -0.4% | Hurt |
| Use OBJ-GSP instead of GES-GSP | Image stitching · DFW | NIQE: 0.4% | Hurt |
| Use OBJ-GSP instead of GES-GSP | Image stitching · REW | NIQE: -2.2% | Hurt |
| Use SAM ViT-B backbone rather than GES-GSP baseline (ablation) | OBJ-GSP · StitchBench | MDR: 1.9% | Helped |
| Use SAM ViT-L backbone rather than GES-GSP baseline (ablation) | OBJ-GSP · StitchBench | MDR: 2.7% | Helped |
| Use SAM ViT-H backbone rather than GES-GSP baseline (ablation) | OBJ-GSP · StitchBench | MDR: 3.6% | Helped |
| Replace SAM with EfficientSAM while retaining mesh sampling (ablation) | OBJ-GSP · Image stitching | stitching quality | No clear effect |
| Replace SAM with EfficientSAM while retaining mesh sampling (ablation) | OBJ-GSP · Image stitching | segmentation time: 5% | Helped |
| Replace object-level triangular mesh sampling with line-based triangular sampling (ablation) | OBJ-GSP · Image stitching | object structure preservation | Hurt |
Surprising findings
- On the SVA dataset, OBJ-GSP's MDR is slightly worse than GES-GSP (1.47813 vs 1.04473), and on the DFW subset OBJ-GSP's NIQE is also slightly worse than UDIS++.
- The method's gains depend on SAM segmentation quality; when SAM fails, OBJ-GSP degrades to GSP-level performance.
Limitations
- Relies on SAM segmentation quality; fails under poor illumination or sparse features, degrading to GSP.
- Higher computational cost than GES-GSP due to large segmentation model.
- Struggles with significant parallax, where it gives no clear overall improvement over GES-GSP.
Experimental setup
Datasets: StitchBench Varied Drone Dataset ICG Drone Dataset LPC dataset SPHP dataset GES-GSP dataset REW dataset AANAP APAP CAVE DFW DHW
Benchmarks: StitchBench
Metrics: Mean Distorted Residuals (MDR) NIQE
Baselines: GSP GES-GSP APAP AANAP UDIS UDIS++ LPC SEAGULL REW SVA SPHP CAVE
Builds on: GSP GES-GSP As-Rigid-As-Possible Segment Anything Model
Closest papers by meaning
- Geometric Structure Preserving Warp for Natural Image Stitching (CVPR 2022 · similarity 0.90)
- Depth-Supervised Fusion Network for Seamless-Free Image Stitching (NeurIPS 2025 · similarity 0.86)
- Reconstructing the Image Stitching Pipeline: Integrating Fusion and Rectangling into a Unified Inpainting Model (NeurIPS 2024 · similarity 0.83)
- Tango3D: Towards Alignment for Global and Local 2D-3D Correspondence (NeurIPS 2026 · similarity 0.83)
- Towards Robust Image Stitching: An Adaptive Resistance Learning against Compatible Attacks (AAAI 2024 · similarity 0.83)
- GeoReF: Geometric Alignment Across Shape Variation for Category-level Object Pose Refinement (CVPR 2024 · 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.