library("rplatypus")
fit <- platypus_fit(spec)
plot_anchors(fit) # the split they were fitted to
plot_anchors(fit, split = "validation") # and one they were notDraw the anchors against the boxes they were fitted to
Description
The picture the 2020 package drew, and the one no summary statistic replaces. Every annotated box is a point - its width against its height, both as fractions of the model’s input - coloured by class, with the anchors on top.
Usage
plot_anchors(object, model = NULL, split = "train", log = FALSE, size = 1.1)
Arguments
object
|
A fit from platypus_fit() on a detection specification.
|
model
|
Which model, when the specification trained several. |
split
|
Which split’s boxes to draw. |
log
|
Draw both axes on a log scale. Detection datasets span a wide range of sizes
|
size
|
Point size for the boxes. |
Details
What it answers that a mean overlap does not: whether a class has any anchor near it at all. A mean of 0.65 can be one class covered well and another not covered at all, and those two situations want different fixes. On blood cells the three classes occupy three distinct regions - median sides of 133, 69 and 26 pixels at a 416 input - and anchors fitted to all three together cover them at 0.88, 0.85 and 0.83, so fitting to the mixture does not abandon the smallest class. COCO’s nine cover the same three at 0.64, 0.70 and 0.70: evenly worse rather than blind to one, which is the useful thing to know. They are not aimed elsewhere - they span 10 to 373 pixels a side because COCO holds objects of every size, and most of that range describes nothing here.
Worth drawing for a split the anchors were not fitted on. Anchors that sit among the training boxes and away from the validation ones say the two halves hold different objects, and no training curve shows that.
Value
A ggplot object.
See Also
detection_anchors() for the numbers, yolo3() for choosing them.