YASAP: Yet Anothe Star Alignment Program

This script uses SURF to align images and the iterative Lucas-Kanade sparse optical flow to finetune the alignment. It assumes multiple shots of the same target with similar views and stacks these shots to denoise.

Dependencies: opencv, numpy, cython

An example workflow

1. Shoot the milky way using high ISO and fast shutter to get pinpoint stars.

2. Convert the raw files to TIFF images in Darktable. Enable lens correction if possible.

3. Create star mask (i.e., a large area that contains most of the sky for alignment) using GIMP.

4. Run YASAP to merge the stack:

   ./yasap.py ../stack1/*  -o ../stack1.tif --mask ../stack1.mask.png \
        --rm-min=1 --rm-max=1
   
   # assuming the camera is stable and thus the images are already aligned,
   # we compute a denoised foreground from the image stack
   ./yasap.py ../stack1/*  -o ../stack1-fg.tif --use-identity \
        --rm-min=1 --rm-max=1
   

5. Merge the two images in GIMP.

Note: use --rm-max and --rm-min to remove outliers including cloud and airplane/satellite trails.

Notes

• YASAP processes the images in a streaming manner and maintains a working memory bounded by constant size (w.r.t. number of input images, but depending on --rm-min and --rm-max settings).
• YASAP uses a numerically stable algorithm to compute the mean image.
• The input images should be given in time order so that neighboring images have similar perspectives. Each image is first matched with the previous image assuming a small displacement. The transformation is then refined against the first image.
• Perhaps counterintuitively at first glance, the correct transformation is homography (eight degrees of freedom) instead of affine transformation (six degrees of freedom) even if we assume an ideal pinhole camera model and infinitely far pinpoint stars (unless you have a spherical sensor).
• For deep sky targets, a rigid transform may be sufficient; in this case, use --use-rigid-transform.
• For deep sky imaging with a long total exposure, consider using --linear-rgb-match to compensate for the sky's brightness change. Also consider using --star-point-quality-thresh to automatically remove images with motion blur.
• Sometimes using star point alignment (instead of optical flow) for refinement gives better results. This mode was added recently, and I am not sure if it is always better than optical flow. Enable it by --refiner star.
• In challenge settings (e.g., images with clouds and heavy light pollution), you can try --skip-coarse-align --remove-bg --refiner star.
• If the inputs are large bodies (e.g., the moon), you can try --refiner optd to use the dense optical flow, with a mask of the body, for better results.
• Use --only-stack --stacker softmax for stacking star trail images with aligned foreground.
• Use ./remove_bg.py to remove sky background.