sky image

README.md

@@ -41,6 +41,7 @@ page from a mid-century journal.
 | Images | 2,001 JPEG, RGB, ~2000–2700 px wide |
 | Model responses | 3 models x 255 figures x 2 conditions = **36,210** answered questions |
 | Ground truth | Generator-exact, not human-annotated |
+| Published figures | **104** real journal figures with AVM astrometry and human-verified plot boxes, for comparison ([below](#published-figures-vqa_realskyimages)) |
 | Download size | ~8 GB of data |
 | License | Apache 2.0 (see [Provenance](#provenance-and-attribution) for the sky cutouts) |
 
@@ -93,6 +94,15 @@ LMM_outputs_n255_archive_light/      RUN 2 -- the SAME 255 figures, aged
   claude_haiku/
   gemini/
 
+VQA_realskyimages/                   104 PUBLISHED figures (metadata only)
+  qa_jsons/   rsky_NNNNNN_qa.json    provenance + AVM astrometry + plot box
+                                     + 4 questions
+  selection_manifest.json            all 163 scraped: kept / excluded, and why
+  fetch_images.py                    re-downloads the images (journal copyright)
+
+LMM_outputs_realskyimages/           Gemini's answers on the published figures
+  gemini/       rsky_NNNNNN_qa.pickle   (score against VQA_realskyimages/qa_jsons)
+
 example_data/                        3 illustrative figures shown twice: as
   imgs/      Picture_N00089.jpeg     rendered, and with every bounding box
   diags/     Picture_N00089.jpeg     drawn on in red, so the annotations can
@@ -108,6 +118,36 @@ numbering and are not part of the 2,001.
 
 ---
 
+## Published figures (`VQA_realskyimages/`)
+
+To test whether results on the synthetic figures carry over to real ones, the
+dataset also includes **104 sky figures from published journal articles**,
+indexed by the AAS Astronomy Image Explorer (AstroExplorer) with Astronomy
+Visualization Metadata (AVM) placing each image on the sky. Their ground truth
+comes from that astrometry, so only geometric questions are asked: angular
+width and height (worded identically to the synthetic set, so scores compare
+directly) and the J2000 position of the field centre.
+
+The figures were verified in two stages before use:
+
+1. **Automated astrometry check (163 → 156).** Each figure's AVM position is
+   compared with the SIMBAD position of the object named in its article title;
+   figures are also flagged for coordinates copied from another article, or for
+   fields too wide for one tangent-plane projection. 7 fail and are excluded.
+2. **Human-verified plot boxes (156 → 104).** The archived images are page
+   crops, so the plotted region is located before its angular size is
+   measured: an OpenCV pass proposes the axis frame, and every proposal is
+   then checked by hand in Label Studio -- accepted, corrected, or deleted.
+   52 are deleted, almost all because the crop cuts off the axis labels.
+
+The reviewed box is stored with every record (`source['plot_box']`), and the
+angular sizes are measured along its edges through the AVM astrometry.
+Images are **not** redistributed (journal copyright); `fetch_images.py`
+rebuilds them. Full details, including the exact exclusion criteria and the
+box fields, are in [`VQA_realskyimages/README.md`](VQA_realskyimages/README.md).
+
+---
+
 ## Citation
 
 <!-- TODO: replace with the WASP 2026 paper reference once it is public. -->

VQA_realskyimages/README.md

@@ -38,22 +38,82 @@ Each json is a JSON-encoded *string* (two decodes, as in `VQA_full`):
 | `VQA` | `VQA[level]['Figure-level questions'][name]` -> entry, same entry fields as `VQA_full` |
 | `vqa_id` | `rsky_NNNNNN` |
 
-## Selection
-
-163 figures were scraped; **7 are excluded** because `astrometry_check` flags
-`marginal_correctness` (wrong position, coordinates copied from a neighbouring
-article, or a field too wide to map with one TAN projection). That leaves 156.
-
-**52 more are excluded after plot-box review, leaving the 104 here.** The
-stored JPEGs are page crops (running heads, colorbars, captions), so the
-plotted region had to be located before its size could serve as ground truth.
-An OpenCV first pass proposed a 4-corner plot (axis) frame for each figure, and
-every proposal was then **human verified** in Label Studio. The reviewer
-deleted the box for 52 figures, almost all because the crop cuts off their
-axis labels, and those figures are dropped. The `rsky_` ids keep their original
-numbers, so there are gaps. The code is in the `real_data/scripts/` working
-area (`detect_plot_box.py`, `apply_plot_boxes.py`, `update_realsky_dataset.py`,
-`label_studio/`).
+## Selection and verification
+
+**163 → 156 → 104.** Every AstroExplorer figure tagged "Images with
+astrometry" was scraped (163). An automated astrometry check removes 7, and a
+human-verified plot box is required of every remaining figure, which removes
+52 more. `selection_manifest.json` records all 163 with the step that dropped
+each, and why. The `rsky_` ids keep their original numbers, so there are gaps.
+
+### 1. Automated astrometry check (163 → 156)
+
+The AVM astrometry was supplied with each figure, not fitted by us, so it is
+checked before it is trusted. Each figure fails the check
+(`astrometry_check.marginal_correctness = true`) if any of these hold:
+
+| Check | Fails when | Excluded |
+|---|---|---|
+| Position | the object named in the article title, resolved through SIMBAD (Sesame), lies further from the image centre than the field's half-diagonal + max(2°, the full field diagonal) | 5 |
+| Copied coordinates | the reference coordinates are identical to those of a figure from a *different* article, and the figure does not match its own title target | 1 |
+| Field size | the field is more than 90° across, too wide for one tangent-plane (TAN) projection | 1 |
+
+The margin in the position check is deliberately generous: extended targets
+(the Cygnus Loop, Orion) can sit partly off-field. Milder problems -- a target
+just outside the field, a 30–90° field -- are recorded as `notes` without
+excluding the figure. Each record carries the result in
+`source['astrometry_check']` (`marginal_correctness`, `reasons`, `notes`,
+`title_target_sep_deg`, ...). As an independent check of the frame, for the
+figures whose title target resolves in SIMBAD the AVM centre lies a median
+0.025° from the J2000 position (see [the frame section](#the-frame-is-j2000-and-it-was-checked)).
+
+### 2. Human-verified plot boxes (156 → 104)
+
+The stored JPEGs are page crops -- running heads, colorbars, captions, sometimes
+neighbouring text -- so the plotted region has to be located before its angular
+size can serve as ground truth.
+
+* **Automated proposal.** An OpenCV first pass proposes the plot (axis) frame
+  as four corners: Otsu threshold; morphological opening with long horizontal
+  and vertical kernels to keep only long straight ink runs; closed contours
+  simplified to near-rectangular quadrilaterals (corner angles 90 ± 12°,
+  covering ≥ 5% of the image); for frameless photographic panels, the largest
+  non-paper region around the AVM reference pixel; and, failing both, a box
+  from the longest horizontal and vertical segments. Candidates containing the
+  AVM reference pixel are preferred, since the astrometric solution is anchored
+  there.
+* **Human verification.** Every one of the 156 proposals was opened in Label
+  Studio (1.23.2) and either accepted, corrected by dragging its corners, or
+  deleted. A box was deleted when no reliable frame could be drawn -- almost
+  always because the crop cuts off the axis labels. **52 were deleted; those
+  figures are excluded**, leaving 104. All 104 kept boxes have
+  `status: "reviewed"`; none is an unreviewed detector proposal.
+
+The box is stored in each record as `source['plot_box']`:
+
+| Field | Meaning |
+|---|---|
+| `corners_xy_topleft` | the four reviewed corners, image pixels of `images_nometa/<vqa_id>.jpeg`, origin top-left |
+| `corner_order` | `"TL, TR, BR, BL"` |
+| `rect_px` | axis-aligned bounding rectangle of the corners `[x0, y0, x1, y1]` |
+| `field_width_arcmin`, `field_height_arcmin` | the box's angular sides (below) -- identical to the `field width` / `field height` answers |
+| `center_radec_deg` | J2000 position of the box centre -- the `center right ascension` / `center declination` answers |
+| `reviewer_flags`, `reviewer_notes` | anything the reviewer recorded (none were used) |
+| `status`, `method` | `reviewed`; `OpenCV first pass (detect_plot_box.py), human verified in Label Studio` |
+
+In `selection_manifest.json` the 52 rejected figures keep a `plot_box` entry
+with `status: "reviewed"`, null corners, and `exclude_reason: "box deleted by
+reviewer"`, so the rejection itself is on record.
+
+**Consistency, checked 2026-10-11:** all 104 records hold a four-corner box,
+and their `field width` / `field height` answers equal the box values.
+Recomputing the widths from the stored corners and AVM parameters reproduces
+them to within 0.013% (`models/figure_notebooks/question_tables.ipynb` in the
+[code repository](https://github.com/jnaiman/SkyImagesWASP2026)).
+
+The selection and plot-box code (`astrometry_check.py`, `detect_plot_box.py`,
+`apply_plot_boxes.py`, `update_realsky_dataset.py`, `label_studio/`) is in the
+`real_data/scripts/` working area.
 
 ## Ground truth comes from the plot box, not the whole image
 
@@ -71,7 +131,7 @@ pixels with the origin top-left, rectangle, reviewer flags/notes) in
 
 ## The questions, and why these
 
-* **`field width`** (Level 2) -- What is the angular width of the sky region shown in this figure, measured along the right ascension axis, as a tru?
+* **`field width`** (Level 2) -- What is the angular width of the sky region shown in this figure, measured along the right ascension axis, as a true angle on the sky (i.e. including the cos(declination) factor)?
 * **`field height`** (Level 2) -- What is the angular height of the sky region shown in this figure, measured along the declination axis?
 * **`center right ascension`** (Level 3) -- What is the right ascension of the centre of this sky image, in the J2000 (FK5) frame?
 * **`center declination`** (Level 3) -- What is the declination of the centre of this sky image, in the J2000 (FK5) frame?