nasa / giant Goto Github PK

The is no method available to build the lost-in-space catalog from the stars catalog.

When the user attempts to resolve the stars on an image without a priori data, it raises the following error:
ValueError: The lost in space catalogue has not been loaded. Cannot solve lost in space problem.See build_lost_in_space_catalogue interface for details.

I couldn't find any script or method to build the lost-in-space catalog.

• Otsu thresholding only returning n-1 thresholds instead of n thresholds (for 4-level otsu, only get 3)
  • In image_processing.py, line 784 (thresholds = unique_iu8[kk[:n - 2]].astype(np.float64)) should be thresholds = unique_iu8[kk[:n - 1]].astype(np.float64)
• Also in otsu, kk should be sorted after being found in sig_fun
  • Between line 775 and 777 in image_processing.py, needs kk = np.sort(kk)
• Outlier rejection does not perform correctly when body is large in frame
  • If the object is greater that ~50% of the image area, the outlier check in identify_subpixel_limbs does not function as expected
  • Starts on line 1899 in image_processing.py, commenting out lines 1899-1909 did not seem to affect any other cases, including when object was small. Should double check necessity of it
• Potential to run out of memory when object is large in frame
  • In _pixel limbs in image_processing.py, if there are too many edge points (typically when object is large in high resolution image), the line edge_distances = np.abs(perpendicular_direction.reshape(1, 1, 2) @ scan_start2edge_points).squeeze() grabs too much memory leading the process to be killed. Seen mainly during close flybys with image sizes of 2048x2048 or larger. Could potentially do piecewise after a check? Binning image and edge pixels to lower resolution? Have not fixed this one yet, interested in hearing your thoughts.

NAIF doesn't support FTP anymore...

First, line 1387 of image_processing.py is as follows:

            standard_deviation = np.nanstd(data[~outliers]) / 2

Should be changed to:

            standard_deviation = np.nanstd(data[~outliers]) / np.sqrt(2)

The rationale is that the difference between two independent normal random variables each with standard deviation σ also follows a normal distribution but with standard deviation sqrt(2)σ. Thus, the standard deviation of the differences is sqrt(2)σ. To estimate σ, you can divide the standard deviation of the differences by sqrt(2).

References: (ours is the case where σ_x = σ_y)

• https://online.stat.psu.edu/stat500/book/export/html/572
• https://mathworld.wolfram.com/NormalDifferenceDistribution.html

Second, line 1351 of image_processing.py is as follows:

            standard_deviation = np.nanstd(flat_dark) / 2

should be changed to:

            standard_deviation = np.nanstd(flat_dark)

In this case, we are taking the standard deviation of the (dark) image data itself, not the differences between points, so it seems like there should be no scale factor. The standard deviation of the data is already directly representative of the noise.

Note that this will effect star ID tuning (poi_threshold), since this is used to compute SNR for each pixel.