OBJ-GSP

Implementation of the AAAI 2025 paper Object-level Geometric Structure Preserving for Natural Image Stitching (Unofficial Port).

This README is rewritten for this fork based on commit history since 98828c91b6db825ec3ec4aad698e03653f319e3a, with emphasis on required engineering changes.

Required Fork Changes (since 98828c9)

1. Migrated from manual .sln management to unified CMake builds.

   • Added top-level and per-module CMakeLists.txt.
   • Split Debugger/Feature/Mesh/Stitching/Util/sam into static libraries and linked them centrally.
   • Why necessary: better maintainability and reproducible Linux/WSL/CI builds.

2. Added CMake presets and vcpkg manifest workflow.

   • Added CMakePresets.json, CMakeUserPresets.json, vcpkg.json, and vcpkg-configuration.json.
   • Why necessary: standardized setup and easier onboarding on new machines.

3. Updated dependency strategy.

   • vcpkg now manages gflags, vlfeat, and eigen3.
   • eigen3 is pinned to 3.4.0 via override.
   • onnxruntime is no longer pulled through manifest and should be provided as a local/system CMake package.
   • Why necessary: improves dependency resolution reliability across platforms.

4. Refactored linking and enabled OpenMP.

   • Added find_package(OpenMP REQUIRED) in root CMake.
   • Submodules now use explicit target_link_libraries dependencies.
   • Why necessary: avoids hidden linker issues and ensures parallel-related paths build correctly.

5. Cross-platform cleanup (removed Windows-specific artifacts).

   • Replaced directory creation with mkdir(path, 0755) in Stitching/Parameter.cpp.
   • Removed VlFeat.props and vl.dll from repository.
   • Removed bundled dirent.h and switched to system header usage.
   • Why necessary: native Linux build compatibility without platform-bound binaries.

6. Header/include cleanup and formatting fixes.

   • Adjusted include paths and Eigen geometry-related headers.
   • Why necessary: reduces compiler-dependent include/symbol issues.

Environment and Dependencies

• Ubuntu 22.04 LTS (container or whatever else)
• CMake >= 3.22
• Ninja
• GCC/G++ (current preset defaults to gcc-9 / g++-9, not guranteed to work with newer versions due to potential C++ grammar changes)
• OpenCV (4.x recommended)
• ONNX Runtime (must be discoverable by find_package(onnxruntime CONFIG REQUIRED))
• vcpkg (for gflags, vlfeat, eigen3)

Build (Recommended)

1. Set vcpkg root:

export VCPKG_ROOT=/root/vcpkg

2. Configure:

cmake --preset vcpkg

3. Build:

cmake --build build -j

4. Run:

./build/main

If CMake cannot find OpenCV or ONNX Runtime, compile and install them manually with the .pc files or CMake config files properly set up.

Datasets

1. StitchBench

   • Sample pair in ./input-data/AANAP-01_skyline.
   • Full dataset: https://huggingface.co/datasets/RussRobin/StitchBench

2. Aerial234

   • Dataset: https://huggingface.co/datasets/RussRobin/Aerial234
   • Related work: https://www.sciencedirect.com/science/article/pii/S1047320323000858

SAM Notes

The repository includes sam/ and sam.ipynb. For SAM-related reproduction, ensure ONNX model paths and runtime I/O expectations match your local setup.

Citation

@inproceedings{cai2025object,
  title={Object-level geometric structure preserving for natural image stitching},
  author={Cai, Wenxiao and Yang, Wankou},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39},
  number={2},
  pages={1926--1934},
  year={2025}
}

@article{cai2023uav,
  title={UAV image stitching by estimating orthograph with RGB cameras},
  author={Cai, Wenxiao and Du, Songlin and Yang, Wankou},
  journal={Journal of Visual Communication and Image Representation},
  volume={94},
  pages={103835},
  year={2023},
  publisher={Elsevier}
}

If you have questions, feel free to open an issue.