Rotate and perspective module: perspective correction skewed with (automatically) resized images

Is there an existing issue for this?

• I checked and did not find my issue in the already reported ones

Describe the bug

In rotate and perspective, the two manual ways of defining structure — "manually define perspective rectangle" and "manually draw structure lines" — give wrong results on images where an earlier module changes the image dimensions (as it is done with Fuji SuperCCD RAF images). Automatic line analysis on the same image is correct.

Steps to reproduce

1. Open a Fuji Super CCD raw file (e.g. FinePix S100FS .RAF). These get rotate pixels applied automatically, which changes the image dimensions.
2. In rotate and perspective, click "manually define perspective rectangle" and place the four handles on the corners of something rectangular in the scene.
3. Click the vertical + horizontal fit button.

The image comes out rotated and skewed.

Further observations

• Every modifier variant of the fit buttons is equally wrong, including the rotation-only fit. A rotation-only fit cannot skew an image, so the line data must be wrong rather than the solver.
• "Automatically analyse line structure in image" on the same image gives a perfect correction.

Expected behavior

the rectangle is corrected to a rectangle.

Logfile | Screenshot | Screencast

No response

Commit

No response

Where did you obtain darktable from?

self compiled

darktable version

5.7.0+762~gefee9282a0

What OS are you using?

Windows

What is the version of your OS?

11

Describe your system

No response

Are you using OpenCL GPU in darktable?

Yes

If yes, what is the GPU card and driver?

No response

Please provide additional context if applicable. You can attach files too, but might need to rename to .txt or .zip

WARNING: The below "Cause" and "Possible direction for a fix" are from Claude AI

Cause

The manual methods record the reference size that their line coordinates are measured against as the pixelpipe's input size — the size before any module has run. The automatic analysis instead uses the module's own input size.

Normally nothing resizes the image before this module, so the two are the same and the value happens to be correct. When rotate pixels is active (for instance with Fuji SuperCCD images from S100FS camera) they differ, and the manual methods then fit against the wrong image geometry: both the wrong aspect ratio and the wrong image center.

Possible direction for a fix

In src/iop/ashift.c, the manual paths take these dimensions from piece->iwidth/iheight, which is the whole pipe's input; the working automatic path uses the module's own input buffer. Taking them from piece->buf_in instead makes the manual methods agree with the automatic one. There are four such places: the rectangle setup, the draw-lines setup, and the two restore-from-saved-parameters paths.

Start in src/iop/ashift.c and inspect the manual rectangle, draw-lines, and saved-parameter restoration paths, comparing their dimensions with the automatic analysis path. Reproduce with a Fuji SuperCCD RAF image whose dimensions were changed earlier in the pipe; done means the manual fit methods produce the same uns skewed rectangle correction as automatic analysis.