opencv_createsamples mixed mode

asked 2018-12-14 03:50:48 -0600

If I read the documentation correctly I can use opencv_createsamples in two distinct modes: either create .vec file from one image applying various distortions (rotations, intensity, etc.) or just convert list of files to .vec file. Is it possible to mix those modes? I mean to give ~15 real annotated samples and apply distortions to all of them to generate eg. 300 samples (20 per input image)? If so, how do I do that? Can I generate separate .vec files for each image and then somehow merge those files? If so, how?

edit retag flag offensive close merge delete

add a comment

1 answer

Sort by » oldest newest most voted

answered 2018-12-14 04:18:51 -0600

berak
32993 ●7 ●81 ●312

while you can try to augment your images (if you do it manually), you probably shouldn't. real life problems need real life data (and a ton of images). any attempts at "faking it" will simply result in a bad classifier.

also note, that the whole cascade based object detection is deprecated, you should use a dnn based classifier instead. latest opencv4 also does no more include the tools to train cascades, or the cuda version of it.