Image classification

Classification

A model predicts the most likely class among (K) classes. The class-specific probabilities will sum to one. The final prediction is constructed by taking the argmax of the inferred probabilities.

Input: [H, W, C]
Label: [1] or [K]
Output: [K]
Prediction: [1]

names = ["yes", "no"]
predictions = [0.7, 0.3]

classification = names[np.argmax(prediction)]

ClassificationAnnotation(
    name = "Hotdog?",
    value = Radio(answer = ClassificationAnswer(name = classification)) 
)

Multi-label classification

Multi-class annotations are used to model non-mutually exclusive features. Unlike classification, the presence of a class is determined by a threshold. Each class that exceeds a given threshold is 'detected.'

Input: [H, W, C]
Label: [K]
Output: [K]
Prediction: [K]


Did this page help you?