Trial: Claude vision can locate accessible parking stalls on ~10 cm/px aerial orthophotos (as suggestions, not counts)
Question: at roughly 10 cm per pixel (1400x1400 px covering 140 m), can a vision model find stalls painted with the wheelchair symbol, well enough to pre-place markers for a human?
Setup: one image per request to claude-opus-5-5, plain messages API, no tiling. Prompt asked for JSON only: a list of stalls with x,y as 0-1 fractions from top-left, kind, the visual cue seen, and a high/medium/low confidence, plus a one-line note on image quality. About 2,800 input tokens per image, 5-11 s each.
Result on 4 commercial lots, checked by drawing the returned points on the image and zooming in: - The model found the correct cluster of accessible stalls on all 4 lots, including symbols inside deep building shadow. - Position error ranged from dead-on to about one stall width (roughly 30 px). Good enough to drop a draggable pin, not good enough for measuring. - Recall was incomplete: one clearly visible symbol in open sunlight was missed while two shadowed ones nearby were found. - False positives were almost all the ones it labelled low confidence: ordinary cars next to yellow hatched no-parking areas, or the striped access aisle itself counted as a stall. - It could not reliably tell van-accessible from standard.
Practical guidance: - Ask for a per-item confidence and the cue it saw. Filtering to high/medium removed most wrong guesses in this sample. - Tell it explicitly that the striped aisle is not a stall, and that yellow hatching alone is not an accessibility cue. - Treat output as suggestions a person confirms. Do not use the count directly for a compliance-style check. - Verify by overlaying the points and cropping at 3x; reading the JSON alone tells you nothing about accuracy.
Not tried: tiling the image into quadrants for more pixels per stall, which is the obvious next step for the position error and the missed symbol. Sample size is 4 images from one city, so treat the hit rates as anecdotal.