Revision history [back]

I see many topics appearing about cascade training for object detection. However, what I do not seem to find are discussions about how fluently training should actually be.

I am training a LBP object detector using the boosted cascade of weak classifiers, done by using the train_cascade algorithm. The detector is used for detecting cars in aerial imagery.

I have used 200 positive examples and 1000 negative examples. I asked to train 35 stages, which the trainer is actually doing, but when I check how fast the algorithm is processing the negative samples, I see weird results.

At first stages processing goes very fast. At stage 19 however, processing goes about a sample each 5 - 10 minutes. Just wondering, since LBP has been reported as beïng trainable in matter of a day, where HAAR needs a week, how fast training goes with others.

Also, I have the idea that assigning more memory to the process, by increasing the precalcValBufSize and precalcIdxBufSize parameter to 2048 MB each, doesn't seem to influence the training more, than actually assigning 500 MB to each. Any suggestions there?

I see many topics appearing about cascade training for object detection. However, what I do not seem to find are discussions about how fluently training should actually be.

I am training a LBP object detector using the boosted cascade of weak classifiers, done by using the train_cascade algorithm. The detector is used for detecting cars in aerial imagery.

I have used 200 positive examples and 1000 negative examples. I asked to train 35 stages, which the trainer is actually doing, but when I check how fast the algorithm is processing the negative samples, I see weird results.

At first stages processing goes very fast. At stage 19 however, processing goes about a sample each 5 - 10 minutes. Just wondering, since LBP has been reported as beïng trainable in matter of a day, where HAAR needs a week, how fast training goes with others.

Also, I have the idea that assigning more memory to the process, by increasing the precalcValBufSize and precalcIdxBufSize parameter to 2048 MB each, doesn't seem to influence the training more, than actually assigning 500 MB to each. Any suggestions there?