Skip to content

arcsec compared with ASTAP

arcsec and ASTAP’s command-line solver, astap_cli, were run on the same 635 test images, given the same position hint, field size and star databases, and scored the same way. The images have known correct solutions, so every answer could be checked, not just counted.

Measured on 12–13 October 2026: arcsec 0.7.0 against astap_cli CLI-2026.07.30, both with the D80, G05 and W08 databases.

596 of the images can be solved; the other 39 are controls (below). A solve counts only if all four corners of the image land within 5″ of the truth (or one pixel, on very coarse images); a solution further out than that counts as a wrong answer.

ASTAP is shown twice: with its default settings, which is how N.I.N.A. and SharpCap run it unless told otherwise, and with -z 1, which turns off its automatic downsampling of large images. That setting is the one change that helped it most.

arcsec ASTAP, defaults ASTAP, -z 1
Solved correctly (of 596) 589 (99 %) 303 (51 %) 339 (57 %)
Wrong answers 1 2 2
Controls answered (of 35 that must fail) 0 0 0

arcsec solved every image that ASTAP solved, at either setting.

Measured on the images both solvers solved, median error in arcseconds:

arcsec ASTAP
Centre, against ASTAP’s defaults (303 images) 0.25″ 0.26″
Corners, against ASTAP’s defaults (303 images) 0.46″ 0.71″
Centre, against ASTAP -z 1 (339 images) 0.25″ 0.26″
Corners, against ASTAP -z 1 (339 images) 0.42″ 0.63″

The two agree at the centre. At the corners, where an error in scale or rotation shows most, arcsec is closer.

Wall-clock time per image, start to finish, one image at a time on a 24-core Linux PC:

arcsec ASTAP, defaults ASTAP, -z 1
Median, images both solve 0.10–0.15 s 0.21 s 0.28 s
Median, images that did not solve 0.32–0.36 s 0.66 s 1.03 s
All 635 images, total 162–196 s 700 s 1285 s

arcsec ran in both runs and its own times differed between them, so the ranges show how much the machine varied. Within each run, arcsec took about half ASTAP’s time on the images both solved, and gave up on an unsolvable image in a third to half the time.

Images n arcsec ASTAP, defaults ASTAP, -z 1
Telescope frames: single exposures from 0.4–2 m telescopes 42 42 41 41
Survey cutouts: sky-survey images cut to a chosen size and position 234 230 131 136
Survey frames: original images from SDSS, Pan-STARRS, Legacy Surveys, TESS and others 255 252 97 128
Simulated camera problems: gradients, hot pixels, colour mosaics, trailing, defocus, clouds, lens distortion 65 65 34 34

On the real telescope frames the two are level. The difference is in the other three sets, which cover a wider range of field sizes and image quality than most people’s own frames.

Long side of the image:

Field n arcsec ASTAP, defaults ASTAP, -z 1
under 0.3° 119 119 39 67
0.3°–0.6° 125 125 97 99
0.6°–1.2° 177 173 111 114
1.2°–2.5° 88 86 52 54
2.5°–6° 43 43 1 2
6°–20° 36 35 3 3
over 20° 8 8 0 0

Most of the 2.5°–6° fields are coarse: 28 of the 43 are TESS images at 20″ per pixel, and most of the rest are 12–13″ per pixel. ASTAP failed nearly all of them with either setting, with G05 and W08 installed.

With no position to start from, arcsec uses its blind index, built from the star database you already have. ASTAP has no index: given -r 180, as N.I.N.A. sends for a blind solve, it searches the sky field by field.

Both were given a position on the opposite side of the sky with -r 180, so neither could use it, and the right field size. The sample was 40 of the images, five from each field-size band, chosen by a fixed rule before the run. arcsec had the default index from D80 (287 MB, for fields of 0.3° and up).

n arcsec ASTAP
Fields of 0.3° and up: solved 30 28 9 (and 2 inexact)
… median time when solved 1.3 s 39 s
Fields under 0.3°: solved 10 0 0 (and 1 wrong answer)
Wrong answers 40 0 1

On the 9 fields both solved, the median time was 1.1 s for arcsec and 39 s for ASTAP. ASTAP’s one wrong answer put a 0.17° field on the wrong part of the sky.

Fields under 0.3° are arcsec’s weak spot in blind solving. The default index stops at 0.3°, so for smaller fields arcsec falls back to searching the sky field by field, which can run for many minutes; ASTAP gives up on them in under a second. A deeper index (arcsec catalog index build --min-fov 0.15, 698 MB from D80) reaches down to 0.15°: with it arcsec solved 2 of the 5 sample fields between 0.15° and 0.3°, in under a second each, and still none below 0.15°. With a position hint, arcsec solves all 119 fields under 0.3° (ASTAP 39, or 67 at -z 1).

  • Wrong focal length. With --fov-search, arcsec still solves when the scale it is given is up to four times too large or too small, for example a forgotten reducer.
  • No scale at all. A blind solve does not need the field size either: without one, arcsec tries pixel scales from 0.3″ to 60″.
  • Image formats. arcsec reads XISF, ASDF, JPEG, PNG and TIFF as well as FITS.
  • Catalogues. arcsec catalog recommend and arcsec catalog install pick and fetch the right star database for your field, and build the blind index from it.
  • Embedding. arcsec is also a C library, which is how a Siril build can call it directly.

What arcsec does not do: ASTAP is also an image viewer, stacker and photometry tool, and arcsec replaces only its command-line solver. The two can share one set of star databases.

  • The images. 635 FITS images with known solutions: 42 telescope frames from Las Cumbres Observatory with their pipeline’s solutions, 489 survey cutouts and frames, 65 simulated camera problems, and 39 controls. The controls are 35 images that must not solve (noise, flat fields, fake star fields, shuffled fields, bright nebula cores, real fields with a hint 30°–60° away) and 4 given a deliberately wrong field size.
  • The same inputs for both. Each solver was given the true centre as its position hint, the image height as -fov, and -r 5. ASTAP chose its own database by field size, as arcsec does.
  • Defaults otherwise. Apart from -z 1 in the second ASTAP run, neither solver was tuned. ASTAP might solve more with other settings for particular images, such as its star count (-s).
  • Scoring. Each solution was checked at the centre and the four corners against the known answer.

The full method, every image and its source, and earlier results are in docs/test-images.md in the repository. To repeat the comparison:

Terminal window
scripts/fetch-corpus.py
scripts/benchmark.py --corpus --auto-db --astap ~/astap_cli --jobs 1