Hex raft, dark board
resolved
- Was
- 1
- off by 18
- Now
- 19
- exact
- Actual
- 19
- ground truth
ValEye · counting engine
What the counter fixed, and what still beats it
Every number here is measured against exact ground truth. The clusters that used to collapse now count correctly; the cases that remain turned out not to be clustering problems at all.
- Photo corpus
- 239/253+1
- Adversarial
- 35/48+14
- Phantoms
- 170−91%
- Overlaps
- 71−64%
Clusters that used to break it
Rafts collapsed to a single pill; textured boards produced hundreds of phantoms.
Hex raft, wood board
resolved
- Was
- 1
- off by 18
- Now
- 19
- exact
- Actual
- 19
- ground truth
37-pill raft, dark
resolved
- Was
- 1
- off by 36
- Now
- 37
- exact
- Actual
- 37
- ground truth
Crossed chains, noise
resolved
- Was
- 261
- off by 248
- Now
- 13
- exact
- Actual
- 13
- ground truth
120-pill dense board
improved only
- Was
- 41
- off by 79
- Now
- 142
- off by 22
- Actual
- 120
- ground truth
Ring (hollow centre)
resolved
- Was
- 1
- off by 11
- Now
- 12
- exact
- Actual
- 12
- ground truth
Shiny beads (real photo)
improved only
- Was
- 49
- off by 15
- Now
- 35
- off by 1
- Actual
- 34
- ground truth
The separation sweep
Real pills cut from four difficult photos — colour, gloss, engraving and shadow intact — laid in a grid whose spacing walks from 3.0× pill width down to 1.7×. Only the arrangement is synthetic. Bar height is the count error; 2.0× is just touching.
r-7ff7fd99matte white caplets on brown
never breaks
s-eb90778fspecular beads (gloss varies by angle)
breaks at 1.9×
r-f5d11815speckled caplets
breaks at 3.0×
lined-bfdbfef9pale pills on ruled paper
breaks at 2.2×
Proximity is not what breaks the counter
Matte caplets survive every spacing, including past contact. Speckled caplets fail at 3.0×, fully separated. Two of the four curves are non-monotonic — failing at 1.9× and passing at 1.7× — which no proximity explanation can produce. The remaining failures are properties of the pill surface and the board, wearing a clustering costume.
Still failing
14 of 253 photos miscount. 9 are off by exactly one pill.
| Photo | Counted | Actual | Off |
|---|---|---|---|
| synth2-cc-light-normal-n60-t25-s298.png | 43 | 48 | -5 |
| lined/lined-69204ff4.jpg | 22 | 20 | +2 |
| lined/lined-bfdbfef9.jpg | 18 | 20 | -2 |
| synth2-rc-wood-small-n12-t65-s229.png | 10 | 12 | -2 |
| lined/lined-503b3041.jpg | 18 | 20 | -2 |
| synth2-rc-light-small-n30-t65-s140.png | 29 | 30 | -1 |
| synth2-cc-kraft-normal-n20-t25-s301.png | 19 | 20 | -1 |
| shiny/s-0bfc44d8.jpg | 33 | 34 | -1 |
| synth2-cc-kraft-normal-n60-t25-s302.png | 59 | 60 | -1 |
| shiny/s-eb90778f.jpg | 35 | 34 | +1 |
| synth2-rc-wood-normal-n12-t0-s233.png | 11 | 12 | -1 |
| real/r-7ff7fd99.jpg | 19 | 20 | -1 |
| synth2-rc-kraft-small-n120-t35-s258.png | 119 | 120 | -1 |
| synth2-rc-wood-normal-n30-t35-s238.png | 31 | 30 | +1 |
The dense board: a fix that works and cannot ship
A greedy overlap peel — drop the deepest-penetrating outline, re-measure, repeat — takes the 120-pill board from 142 to 124 with every real pill still found. A fixed threshold cannot do this, because real touching pills overlap as much as 24.68px. But scoped or unscoped it costs corpus images, and the two it breaks are dense synthetics whose pills are legitimately packed: by every measure available, a genuinely dense board and one full of invented outlines look identical.
The shiny beads: root cause found, fix unaffordable
Five pills are discarded by a rule that drops blobs below 70% of median brightness as surface rather than product. That assumes one medication means one colour — true for matte tablets. A glossy pill is a small mirror, so its brightness comes from the angle it lies at. The discarded blobs are 2773px and 2064px against a median pill of 1435. Exempting pill-sized blobs costs seven corpus images; adding a length test still costs three, because dark surface marks are frequently both pill-sized and pill-long.