Nitpick

Level

A camera on a motorbike leans. The Mac finds the straight edges (poles, walls, window frames), works out which way is up, and turns the sphere upright, so every later cut stands straight. The camera leaned 44.1° in this frame.

Tiles

Each sphere is cut into 24 views: eight directions, three heights, each 50° wide and 30° tall at full camera size. Small signs stay the size they were shot.

The Mac reads

Apple Vision, on the Mac itself, reads Thai and English on every tile, finds faces (pixelated before any picture leaves the machine) and names the scene: mural, lantern, parked motorbikes. 288 tiles, 359 pieces of text on this set. Green = sure, orange = likely, red = a guess.

1. งราย (0.5)
2. MAISSEE (0.5)
3. บัญได้ (0.5)
4. go tresh (0.5)
5. No GisตีRoti oece ta (0.3)
6. Ce (0.5)

Scene labels on this tile: outdoor 0.98, sky 0.98, cloudy 0.94, blue_sky 0.89, cord 0.63

Settled, or cut out

A read that names a Mot Dang listing within 60 metres, with every other line read sure, is settled: no picture needed. Everything else becomes a crop, enlarged until each line of text is 26 pixels tall, and packed onto a sheet. This set: 3 settled, 72 sign crops, 17 scene crops.

The start of the digest the model reads first

m000 GSAB2131_145..GSAB2131_145 (1) lead GSAB2131_145 14:01:44 18.763627,98.992999 sweep w001.jpg
  near: cm-curated-paper-41945 Dezpax  เดซแพค  เชียงใหม่ แพคเกจจิ้งพิมพ์โลโก้ | cm-curated-box-shop บ็อกซ์ ช็อป | cm-osm-node-3810861068 มิ่งมิตร
  c0067 GSAB2131_145 y-126 p16 t001.jpg :: เก้อไหมใน(0.30) / ที่ดับ(0.50)
  c0068 GSAB2131_145 y-124 p6 t001.jpg :: ดานคร(0.50) / L E(0.30) / ตรงจน(0.50) / นโอก ลิสชิ้ง จำกัด(0.50) / รั้นจัดไฟแม่(0.50) / สามาแล้ง(0.30) / AIRA LAB(0.50) / 053-27E(1.00) / 202(0.50)
  l0077 GSAB2131_145 y90 p3 graffiti 0.26 l001.jpg
  c0038 GSAB2131_145 y112 p3 t001.jpg :: รบ(0.50)
  l0076 GSAB2131_145 y135 p-22 clothing 0.82 l001.jpg

Sweep

Each spot also goes out as one low-resolution 360 strip. It shows where to look, not what a sign says, so the model can catch what no detector flagged: spirit houses, murals, wrapped transformer boxes, bikes parked in a row.

Zoom

When a crop leaves a doubt, the model asks for that exact direction at camera size, and can ask for the frames either side to pick the sharper one. The reader took 14 zooms on this set.

One reader, few turns

A model re-reads everything in its context on every turn, so turns cost more than pictures. The reader gets one batch: the digest, then every page at once, then every zoom at once, then one write. This run took 9 turns; the old reader of these frames took 138.