ImageSticher

A small, focused panorama stitcher for three horizontally-overlapping images (left, center, right).

Problems this project solves

• Left or right images being cut off after warping: the canvas size is computed from the transformed corners of all images so no content is clipped.
• Stitched results ending up in the top-left with large black areas: images are translated into a single positive-coordinate canvas so content appears centered rather than stuck to (0,0).
• Poor blending that only affected one side or left seams visible: blending uses distance-transform weighted blending (feathering) so seams from left/center/right are smooth and the center image is naturally prioritized.
• Smoothing did not affect matching: smoothing is applied only to the copies used for feature detection, so --smooth improves matching robustness without blurring the final output.

How it works (brief)

• Detect features using SIFT on optionally-smoothed copies of the input images.
• Match features (Lowe ratio test) and compute homographies mapping left→center and right→center.
• Compute bounding box of all warped corners to create a canvas large enough to hold everything.
• Warp the original (non-blurred) images into that canvas and composite:
  • With blending (default): use distance-transform weights and light Gaussian blur to compute normalized per-pixel weights, then blend left, center, right.
  • With --no_blend: compositing order is center → left → right (later images overwrite earlier pixels where present).

Requirements

• Python 3.8+
• numpy
• OpenCV with SIFT support (install opencv-contrib-python):

pip install numpy opencv-contrib-python

If you get ModuleNotFoundError: No module named 'cv2', install the packages above.

Usage

From the repository root run:

cd src\test_images
python ..\main.py left.jpg center.jpg right.jpg        # blend (default)
python ..\main.py left.jpg center.jpg right.jpg --smooth    # smooth for matching
python ..\main.py left.jpg center.jpg right.jpg --no_blend  # no blending, overwrite order

• Output: panorama_result.jpg written to the repository root.
• Flags:
  • --smooth — apply Gaussian blur to images used for feature detection (improves matches on noisy images). Final output uses original images.
  • --no_blend — disable feathered blending; images are composited center → left → right with overwrites.

Files

• src/main.py — the stitcher implementation and CLI entrypoint.
• src/test_images/left.jpg, src/test_images/center.jpg, src/test_images/right.jpg — sample inputs included for testing.

Examples

Input images (left, center, right) shown side-by-side:

Left	Center	Right

Stitched panorama result:

If the panorama doesn't look right (alignment or seam issues), try:

• Running with --smooth to improve feature matching.
• Ensuring there's sufficient overlap(~30-40%) between images (SIFT needs common visual features).
• Inspecting the console output to confirm homographies were found (the script falls back to pairwise stitching if matches are insufficient).

Notes

• The code favors preserving final-image sharpness by only smoothing for detection.
• SIFT requires opencv-contrib-python on most installations — use that if SIFT is unavailable.

---

contributions and suggestions are welcome! This is a personal project to explore panorama stitching concepts, so it's not production.