Perspective Correction

^{Kastner Mansion, Newark — Wikimedia Commons, Djflem, CC0. Pitch +12.7°, roll +2.4°.}

Straightens architectural photographs. Two modes:

• Auto: converging verticals made plumb and the frame levelled: roll, pitch and focal length, a camera rotation H = K R K⁻¹, so it cannot shear.
• Horizontal auto: an obliquely shot facade turned into a frontal view, with verticals plumb and horizontals level, so the wall is a true rectangle.

You review one photograph at a time and nothing is written until Save.

Horizontal auto: frontal facades from oblique shots

^{Mariánské Hory, panelové domy — Wikimedia Commons, Lukáš Mižoch, CC BY-SA 2.5. Yaw 59.4°, facade strip on the pink wall, lama fill.}

This mode is for facade textures for 3D, Google Street View grabs, elevation tracing, or any wall you could only photograph from the side. The receding direction is undone (yaw), and the stretch that undoing causes is compressed back to plausible proportions. On a corner view only one wall can come out frontal: drag the facade strip onto that wall. A drawn horizontal line overrules the detector.

Detectors first, a few seconds of your hand second

A line detector (lsd, or mlsd, which keeps only structural edges) finds the facade, and on most photographs that is the whole job. Where it is fooled by cars, a sloping street, roof battens or a second wall, a few seconds of hand work on the picture fix it:

• paint it out (b), box it in with SAM (s), or name it (Grounding DINO);
• strike slanted lines in one click;
• draw the truth (m): a drawn line comes out exactly vertical or horizontal, and 2 + 2 drawn lines make a rectangle with 90° corners.

Every control explains itself on hover (after a second).

Install

pip install -r requirements.txt        run it from this folder
pip install -e .                       or install it, and get the `pc` command
pip install --no-deps ai-edge-litert   optional: the mlsd detector
pip install --no-deps simple-lama-inpainting   optional: lama fill (default; telea without it)

Python 3.9+, numpy, OpenCV, Pillow, piexif. The GUI needs Tkinter. --no-deps matters: lama's pins otherwise break OpenCV. python rectify.py --doctor says what is present. Segmentation weights (BiRefNet, Grounding DINO, SAM 2) are not in the repo; see docs/masking.md.

Developed and tested on Windows only. Linux and macOS are untested.

Use

python rectify.py --gui                          the review window
python rectify.py "D:\Fotos"                     batch: write Foto_corr.jpg beside each
python rectify.py "D:\Fotos" --horizontal        batch, Horizontal auto
python rectify.py --help                         every option

or double-click Perspective Correction.bat. The biggest accuracy win is telling it the lens: --focal-35mm 24.

More: accuracy · algorithm · detectors · masking.

Credits and licence

MIT. See LICENSE and CREDITS. Descended from chsasank/Image-Rectification (review); design informed by darktable's ashift and ShiftN (prior art), no code taken from either. README photographs from Wikimedia Commons, credited beneath each; the test pool is listed in tests/assets/README.md.