[PierreRaybaut]

Fixes from a review of Sigima 1.4.0 before its release (one commit per issue).

Summary

Analysis results of non-uniform images (95cb972)

compute_geometry_from_obj converted the pixel indices returned by the image analyses with dx, dy, x0 and y0, whatever the coordinates of the image. These values do not describe images with non-uniform coordinates, so the results of centroid, enclosing_circle, peak_detection, contour_shape, blob_dog, blob_doh, blob_log, blob_opencv and hough_circle_peaks were:

• given in pixel indices (default dx = dy = 1, x0 = y0 = 0). Example: the centroid of a blob at pixel (51, 51) was (51, 51) instead of (58.4, 57.6);
• or computed with a stale uniform calibration left over by switch_coords_to("non-uniform"). Example: 35.5 instead of 72.5.

For non-uniform coordinates, the indices (offset by the first pixel of the ROI, in index space) are now converted with axis_to_physical. Radii and semi-axes are scaled by the local pixel size at the shape center. The conversion of images with uniform coordinates is unchanged (bit-identical). The nine contracts are at version 2.

ROIs holding no pixel (fdfcf28)

extract_roi and extract_rois silently created an empty image when the regions of interest held no pixel of the image. For example, 5 of the 9 cells of the default ROI grid on a 4×4 image hold no pixel. A rectangle outside the image or with no width does the same, and so does a circle passing between pixel centers. They now raise an explicit ValueError. Inputs that were accepted are now refused, so both contracts are at version 2. The mask of extract_rois is computed once. generate_image_grid_roi is unchanged: the DataLab grid editor relies on it.

Radial profile of an empty image (3481b82)

get_radial_profile failed on empty data with an obscure NumPy error (zero-size array to reduction operation maximum). It now raises an explicit ValueError. The contract is unchanged, since these inputs were already refused.

Dependency documentation (582d274)

requirements.txt and doc/requirements.rst still listed guidata >= 3.16.0 after 8c51eaa. They are regenerated with guidata.utils.genreqs, and the French catalogs follow. The release notes now state the guidata 3.17.0 requirement.

Release notes

• New bug-fix entries, with their French translations.
• The "Operation contract versions" entry lists the contracts increased after the 1.4 exception.

Validation

• New geometry_nonuniform_unit_test.py (16 tests):

  • centroid and peaks on irregular coordinates, with and without ROI;
  • stale calibration;
  • local scaling of radii and semi-axes;
  • uniform vs. equivalent non-uniform results for every shape kind, with NaN padding kept;
  • exact regression of the uniform conversion.

  The non-uniform tests fail on develop.

• New tests for the extraction of ROIs holding no pixel (uniform and non-uniform, grid cells) and for the radial profile of empty images. test_empty_image_roi, which expected an empty image, now expects the refusal.

• Full suite: 2389 passed. pytest --qualification-report: no problems; the 11 increased contracts are qualified at version 2 with exact replays.

Follow-up

DataLab-Platform/DataLab: branch fix/release-1.4-review (CI uses this Sigima branch). It regenerates the Desktop families reference file and adds the release notes.

The Web families reference file of DataLab still records centroid and peak_detection in version 1: these activities are refused, as expected, until DataLab-Web regenerates it.