SeedCounter in full
The home page shows the essentials. Here is the rest: each feature explained as an animation and shown in the app, why it runs in the browser, the eye test and what we have already checked against the literature.
What it does
On the left, the animation that explains what happens. On the right, the same step in SeedCounter. When the real recording already lives on another page, the link takes you there.
The seed is separated from the background by color and the outline is measured on the spot.
Otsu threshold, outline and Feret
On screen, in the app
The real recording, version 4.1: ten faba beans, one click on each, and the count climbs to 10. It is in the faba bean example, with the scale and the check against the authors.
See the recording in the exampleSeeds stack up by length and the seed lot becomes a distribution.
50 model seeds
A ruler photographed next to the seeds becomes the scale. The file's DPI is only a declaration.
10 mm = 1,905.8 px · 1 px = 5.25 µm
On screen, in the app
The real recording, version 4.0: dragging from one centimeter mark to the next measures 1,885.5 px and the scale comes out at 5.304 µm/px. It is in the Manual, in the calibration task.
See the recording in the ManualArea, circularity, solidity and the Feret diameters of each seed, with the axes drawn on it.
The numbers of a shapethe rotating caliper
On screen, in the app
The real recording, version 4.1: the largest of the 10 faba beans is 21.5 × 13.4 mm and 228 mm², with the measurement axes drawn on each one. It is in the Manual, in the measuring task.
See the recording in the ManualViable and non-viable seeds get marks of different shapes, a disc and a ring, so the figure stays legible even printed in black and white.
tetrazolium, 36 of 65 viable
On screen, in the app
The real recording, version 4.0: 36 viable seeds with a cyan disc and 29 non-viable with a magenta ring, as in the group's ground truth. It is in the orchid example.
See the recording in the exampleAnd also
Counts day by day and at each withdrawal from storage, comparing treatments and seed lots.
Storage and viabilitySeveral images in a queue, multipage TIFF and a scan sliced into plates.
For everyday lab workCapture straight from the camera connected to the computer, without going through another program.
The right benchProcessing happens in the browser. No image goes to a server.
Why it runs in the browserCome back to a plate later, with the image, the marks and the outlines.
Open SeedCounterTest report as PDF, CSV, SQL and a YOLO dataset to train your own model.
Exporting, in the ManualTest your eye
The plate appears for a few seconds. Make your guess and watch SeedCounter count them one by one.
More animations in LearnThe plate stays on screen for 4 seconds and then disappears.
Plates generated on the spot, with model seeds. In the app, you count your own image: open SeedCounter.
Science
Before it becomes a feature, each measurement is checked against published data. What we still get wrong is written down.
On 1,904 rice grains measured by SeedCounter and in the Koklu et al. dataset, seven descriptors matched with 0.00% error. The other differences come from calculation conventions.
The rice exampleTheory predicts 8(√2−1)/π = 1.054786. The published dataset shows 1.054052. A difference of 0.07%.
Shape and geometryThe outline simplified to 48 sides overestimates circularity in seeds with low solidity. That is measured and explained, and the fix is the next step.
For labs
With raking light, a seed's shadow touches its neighbor and two become one. If you're going to use it every day, we work with you on what goes around the app: the capture bench, the classes in your protocol and validation with your own samples.
Where it came from
SeedCounter began as an answer to a bench problem at GPEOrq and GPSEM, the orchid and seed research groups: how to count and measure orchid seeds, among the smallest there are, without spending the whole day at the microscope. The group had been reading tetrazolium tests from scanner images since 2011; the app turned that reading into counts, measurements and a test report.
The app comes with example scenes of known count, so you can compare what it finds.
Usage analytics
With it, Google Analytics and Microsoft Clarity show which pages help and where people get lost. No ads, and nothing loads before you choose.