1 A leaf photo
Drop a leaf photo
or click to choose one · or use your camera.
The image is read in this tab and never uploaded.
The image is read in this tab and never uploaded.
Or click a sample
The first is a PlantVillage leaf, the kind the model was trained on. The rest are real field photographs from PlantDoc that it has never seen — including the same disease as the first. Comparing those two is the whole point of this page.
Input
preprocess
— ms
inference
— ms
FLOPs
—
2 Diagnosis
3 What this model is actually worth
—
PlantVillage
detached leaf, grey background, studio light
detached leaf, grey background, studio light
—
PlantDoc
real photographs, soil, hands, sky
real photographs, soil, hands, sky
The first number is the one papers quote, and it is the one that means less.
Every PlantVillage image is a single detached leaf photographed against a uniform
background, so a network can score in the high nineties partly by reading the backdrop.
Evaluated on PlantDoc — the same diseases, photographed in real fields, never seen during
training — the same weights score far lower. Both numbers are measured in the training
notebook, on the classes the two datasets share, with the decision restricted to those
classes so the comparison is fair.
So treat this as a demonstration, not an agronomist. The useful engineering conclusion is that a field-ready model needs field data, and no amount of extra backbone fixes a dataset whose backgrounds carry the label.
So treat this as a demonstration, not an agronomist. The useful engineering conclusion is that a field-ready model needs field data, and no amount of extra backbone fixes a dataset whose backgrounds carry the label.
Runs entirely in your browser. 62k parameters, int8 weights with one scale per
output channel, hand-written convolutions on typed arrays. No upload, no server, no API
key. The Tchebichef arm applies the transform per colour channel and feeds the network a
K×K coefficient map instead of the pixels.