A few observations on QA consistency and image suitability
#2
by leonChen - opened
Thanks for sharing X-PCR. While inspecting the downloaded dataset for ophthalmic VQA, I noticed a few points that might benefit from clarification or future updates:
- Conditional questions and “not applicable” options. Some images receive “No” for disease presence but are still assigned a disease stage or subtype. For example, the DR severity question lacks a “No DR” option. Adding conditional filtering or an N/A answer could improve consistency.
- Matching questions to the actual image type. Some OIMHS inputs contain segmentation overlays, while some Harvard-GDP inputs are RNFL thickness maps paired with B-scan-oriented questions. Separating these input types and checking question applicability could make evaluation clearer.
- Answer distributions and provenance. Several templates have highly concentrated answer distributions. Publishing per-template distributions, majority-answer baselines, and links to supporting source annotations would make the dataset easier to interpret and reuse.
These observations come from consistency checks on the downloaded release, rather than a full expert re-annotation.