OpenCV 2.4 — GPU‑Accelerated Real‑Time Image Stitching

This repository extends OpenCV 2.4 with a GPU‑optimized stitching pipeline centered around cv::CachedStitcher. Calibration (features, matching, bundle adjustment) is performed once; per‑frame composition runs on the GPU using cached warp maps, seam‑aware feather weights, per‑channel exposure compensation, and ROI‑based blending. A complete CPU fallback is available.

Architecture

flowchart LR
  subgraph "Calibration (once)"
    A["Features + Matching"] --> B["Bundle Adjust"]
    B --> C["Build Warp Maps (GPU)"]
    B --> D["Warp Preview (CPU, seam scale)"]
    D --> E["Seam Find (GraphCut)"]
    E --> F["Feather Weights (GPU)"]
    D --> G["Per-Channel Exposure Gains (tile grid)"]
  end

  subgraph "Runtime (each frame)"
    U["Upload (pinned, async)"] --> W["Warp (GPU, cached maps)"]
    W --> X["Exposure (CUDA, per-tile RGB)"]
    X --> Y["Blend (GPU, ROI + weights, normalize)"]
    Y --> Z[Download]
  end

  C -->|cached xmap/ymap + ROI| W
  F -->|weight maps| Y
  G -->|gain grids| X

Loading

Key Features (implemented)

• Cached transforms and ROI placement per image; panorama sized to exact extents.
• GPU warping via cached maps; no in‑place ops; multi‑stream capable.
• Seam‑aware feather weights built on GPU from seam masks.
• Per‑channel exposure compensation using a smoothed 32×32 tile grid (CUDA).
• ROI‑based blending with per‑pixel normalization; CPU fallback to OpenCV Stitcher.
• Optional pinned uploads; per‑stage timings (upload/warp/exposure/blend/download).

Build

• Requirements: OpenCV 2.4 toolchain; CUDA Toolkit (for GPU); CMake ≥ 2.8.12.2; C++11 compiler.
• Configure:
  • mkdir build && cd build
  • cmake .. -DCMAKE_BUILD_TYPE=Release -DWITH_CUDA=ON -DCACHED_STITCHER_BUILD_BENCHMARKS=ON
  • cmake --build . --parallel
• Binaries:
  • realtime_stitching (runtime --gpu/--no-gpu)
  • realtime_stitching_gpu, realtime_stitching_cpu (defaults baked in)
  • benchmark_stitching, benchmark_stitching_gpu, benchmark_stitching_cpu

Usage

• Images: ./realtime_stitching_gpu img1.jpg img2.jpg ...
• Videos: ./realtime_stitching_gpu --video cam1.mp4 cam2.mp4
• Webcams: ./realtime_stitching_gpu --camera 0 1
• Benchmark: ./benchmark_stitching_gpu --iters 300 img1.jpg img2.jpg
• Flags: --bench prints per‑stage timings; --no-gpu forces CPU path.

Project Structure

• include/CachedStitcher.hpp public API, cache layout, performance stats.
• src/CachedStitcher.cpp calibration, ROI/weights/gains, GPU composition.
• src/gpu_transform_cache.cu CUDA kernels (warp, exposure, blending helpers).
• examples/realtime_stitching.cpp demo with optional bench output.
• benchmarks/benchmark_stitching.cpp headless timing tool.
• CMakeLists_CachedStitcher.txt library + binaries; CPU/GPU split executables.
• modules/ OpenCV 2.4 sources (warpers, seam finder, blender, etc.).

Notes

• Seam finding runs at seam scale via the OpenCV API; weights are then generated on GPU.
• Performance depends on content, resolution, GPU, and I/O. Use the benchmark tool to measure on your hardware.