Make detectMultiScale return number of neighbors joined into each object (Feature #3067)

While numNeighbors is a decent measure of confidence, theoretically, the levelWeights provided by the undocumented call to detectMultiScale seems to be a better measure. Assuming issue #3064 is fixed, levelWeights will return the final calculated "feature value" (as calculated from here) that OpenCV internally compares against the last stage's threshold (see here).

This excerpt , for example, describes the measure of confidence as the sum of the "feature values" minus stage thresholds at each stage. OpenCV could be modified to allow this (and in fact, I've tried it here), but the feature value at the last stage is a decent substitute.

Do note that this is not to say that the number of neighbors is not useful, but it should be noted (in the documentation) that it is not the main measure of confidence.

Thanks for the link, it's an interesting research. It would be interesting to try continuous cascade responses rather than binary. Although, currently this isn't supported so I don't think it's worth mentioning in the docs as they describe what the function can already do.

Here's the pull request: https://github.com/Itseez/opencv/pull/961

The old pull request was reverted (see discussion in its comments), here's a new one: https://github.com/Itseez/opencv/pull/976.
This issue's assigned version should be set to 3.0 again.