CODEWITHSUNDEEP
machine-learningprobabilityregressionfeature-selection
Eric
Eric

Reputation: 2341

Supervised Learning

The question is about combining output probabilities from detectors. Each of my detector d_i gives me a probability p_i of object presence in the scene. I have a labelled data set and I would like to learn the weighted combination of the detectors to get a (more accurate) global detector.

What algorithms can I use for such problem? I would like to compare different methods. As a first step, weights could be fixed learnt values. Then it would be nice to make them adaptive according to a time-context for example.

Thanks for ideas.

Upvotes: 0

Views: 61

Answers (1)

bogatron
bogatron

Reputation: 19169

A good place to start is the weighted majority algorithm, which develops the weights for your detectors, based on their relative performance over a data set. There are variations of this algorithm, some of which can handle targets that drift over time.

If your detectors are all instances of the same model (with different parameters), you could also take a look at the various ensemble learning algorithms, although these may be less applicable if your interest is in comparing the relative performance of your base algorithms.

Upvotes: 1

Related Questions