“For simplicity, and since you don’t know much biology and human anatomy, I will tell you how I was built as if I were a building.”
Milestones 2 and 3, Group 7: from 7,343 raw files to a detector for 73 everyday objects in any layout, explained object by object. Twelve stations, left to right. Flip any card for the full site report.
1 / 10←→ move · Enter flip
PUPIL
correctbox offwrong IDextramissedtrue box
allerrors
STRESS TEST
Runs many random grids of held-out test photos with the settings on the eyebrow (grid, mix, conditions, model, threshold) and reports the totals. Grids use seeds 1000, 1001, ... so a run can be repeated exactly.
GRIDS100
FILES
The last detection, in the milestone's output format, plus a snapshot image for reports and slides.
ABOUT
The detector reports the Object ID and bounding box of every object. Random grids and the stress test use only held-out test photos whose true boxes were checked by a person.
How to use: ① set up the test on the eyebrow · ② press RANDOM (or upload) and DETECT · ③ review below · ④ press COMPOUND in the left dock to break the pupil into every model's pupil. Click the pupil for the large view; click an error to zoom onto it.
result
meaning
correct
right Object ID and the box overlaps the true box by at least half (IoU >= 0.5)
box off
right Object ID, but the box overlaps by less than half
wrong ID
box in the right place, another object named (look-alikes flagged)
missed
no box; the reason says whether it was seen below the threshold or is very small
extra
a box matching no object
THE EXTENSION
After Milestone 2, every usable box was checked by a person (two passes, near-duplicate copies removed), and the detector was fine-tuned again on
rotated grids, photos scattered on real backgrounds and objects cut out with SAM, at a higher input resolution. The layout chips in the Mix panel
show these test scenes; their objects come from held-out test photos on unseen backgrounds. Each model keeps its own threshold chosen on validation,
one for grids and one for scenes; drop any image onto the eye to try it.
BEFORE VS AFTER
COMPOUND VIEW (v3.4)
The compound switch in the left dock shows how every model does on the same image: boost, boost2, the Milestone 2 model and a YOLOv8-s, plus the Milestone 1 classifier on each grid cell when its weights are present. On images with true boxes every model is judged against the truth, like the answer, and shown as a score ring and a row per object; on uploads the other models are compared with the answer's boxes. Press compound after a detection and the pupil breaks into one pupil per model, each ringed by its score, and panel 04 opens on its compound tab: hover a box to follow that object through every pupil, click a pupil to open it. Press compound again (or C, Esc, RANDOM, upload or DETECT) to return to the single pupil. It explains; it never changes the answer or the score. Combining the models into one answer was tested on validation under a sealed rule and did not pass (better grids, worse cut-out scenes and more boxes on empty backgrounds), so it stays a view.
MODELS
BY INPUT TYPE (MILESTONE 2 MODEL)
The iris is the score: one segment per object, coloured by its result. After DETECT the view moves into the pupil to show the result; Esc or a click returns. Keys: N random grid · D detect · U upload · R replay the opening · Esc close or zoom out.
Online EYE · upload or drag in an image, then DETECT · random grids, scenes and the stress test
use the course photos and run only in the local EYE