Seems to be a bug in the training code. For training I use 100 positive images and 300 identical negative images. I use 1 stage for training to make it simple and very low FA. The stage gets constructed with 13 classifiers and it reaches a very low FA.
But if I use the same negative images for detection I get 26 falses (neighbor set to 0). That means that the estimate of the falsing rate in the training code has an issue. I use the same scaling factor as used by the training code = 1.41.
Any ideas. I am checking the source code right now.
PARAMETERS:
cascadeDirName: classifier
vecFileName: samples.vec
bgFileName: negatives.txt
numPos: 90
numNeg: 300
numStages: 1
precalcValBufSize[Mb] : 1024
precalcIdxBufSize[Mb] : 1024
stageType: BOOST
featureType: HAAR
sampleWidth: 20
sampleHeight: 20
boostType: GAB
minHitRate: 0.999
maxFalseAlarmRate: 1e-005
weightTrimRate: 0.95
maxDepth: 1
maxWeakCount: 100
mode: ALL
===== TRAINING 0-stage =====
Training until now has taken 0 days 0 hours 0 minutes 13 seconds.
↧