flag todo

VQA_realskyimages/README.md

@@ -79,6 +79,38 @@ Their wording is **identical to the synthetic benchmark's**, so scores on
 `field width`, `field height` and `pixel scale` are directly comparable between
 real and synthetic figures.
 
+## TODO: the positional questions have no synthetic counterpart
+
+`center right ascension` and `center declination` cannot currently be compared
+against anything in `VQA_full`. The synthetic sky figures ask for **axis
+limits** (`minimum/maximum right ascension axis limit`, and the declination
+pair), which presumes a drawn, labelled axis -- fine for a figure we rendered,
+but not something every published figure has. So the real-sky set asks for the
+**centre of the image** instead, and the two are not the same question.
+
+The consequence: the three geometry questions (`field width`, `field height`,
+`pixel scale`) are worded identically in both sets and can be read side by
+side, but the positional result here -- median 0.45 deg, 70% within 1 deg on
+156 figures -- is a standalone number with no synthetic baseline to sit
+against. There is no way to say whether that is good *relative to* performance
+on figures where we know the answer is cleanly readable.
+
+Worth closing later, if the comparison turns out to matter. Either direction
+would work:
+
+* **Add centre questions to the synthetic set.** The generator knows each
+  figure's central RA/Dec exactly, so `center right ascension` /
+  `center declination` could be emitted for the synthetic sky panels using
+  the wording already used here. Cheapest option, and it makes the two sets
+  line up on all five questions. Costs a re-run of the synthetic sky figures.
+* **Add axis-limit questions to the real set.** Harder and probably not worth
+  it: the AVM header gives the image footprint, not what the figure's axes
+  are annotated to, and many of these figures have no usable axis at all.
+
+The first is the better route. Note that if the synthetic centre questions are
+added, they should carry the same explicit "J2000 (FK5)" framing used here, or
+the two still will not be measuring quite the same thing.
+
 ## Licence
 
 The questions, ground truth and metadata in this directory are released under