\n
OpenCV’s standart cascades allow to detect faces and eyes.\u00a0 I wanted to create cascade in similar way to detect another objects:\u00a0 pringles or plate for example.\u00a0 I found some material in Net how to use OpenCV training tools, also I investigated training tool’s source\u00a0 code myself\u00a0 to found out, what training parameters can be tuned.<\/p>\n
For training, I needed thousands of images, containing my object with different lightning conditions and perspectives. After trying to find required number of pictures with Google ,\u00a0 I understood, that it’s really difficult task =).\u00a0 So I decided to take video with my object<\/span>,\u00a0 then I wrote simple program to crop object from video, frame by frame.\u00a0 In such way,\u00a0 I generated about 3000 positive samples (cropped images\u00a0 with my object).\u00a0 Resolution varied from 50×50 to 100×100.\u00a0 The advantage of this method – you\u00a0 get many samples with different reflections, illuminations and backgrounds.\u00a0 It’s very important, that all these images “features” are various!<\/p>\n <\/p>\n Negative samples I got\u00a0 from Internet public databases…, there were about 5000 images of different resolution, which did not contain my objects.\u00a0\u00a0 I worked with BMP format, due to compression artefacts in some compression levels of JPEG.<\/p>\n So I split all my images into 4 sets:<\/p>\n Of course, I could use the same images for testing and training, but it’s not good.\u00a0 Positive images should be be packed to OpenCV vec-file.\u00a0 it can be done with createsamples program from OpenCV.\u00a0 When packed, all images are resized to 20×20 – it’s good size for training speed.\u00a0 Also during object detection, object with smaller size will not be found.\u00a0\u00a0 For negative samples,\u00a0 I only created negative.dat(list of negative file names) with win dir \\b command.<\/p>\n OpenCV implements Adaboost algorithm in haartraining. exe.\u00a0 This tool\u00a0 generates xml cascade file, which is used in haardetection tool for object detection.\u00a0 There are some main characteristics of cascade:\u00a0 number of stages, type of classifiers on each stage.\u00a0These parameters can be configured when you launch haartraining. I configured it as per my training data (positives and negatives).<\/p>\n \n Training is finished, when required false alarm of cascade is achieved.\u00a0 So cascade can contain less stages than nstage – it’s ok. It’s rather worse, when we achieve required number of stage, but required false alarm is not. It’s signal, that something is wrong either with content or with haartraining parameters.\u00a0 Also maxfalsealarm is false alarm of stage, false alarm of cascade is multiplication of stage’s alarms.<\/p>\n I used P4 2.7GHZ with 2 GB and it took about !6 days! of permanent working to finally get cascade.\u00a0 If you have Hyper Threading, you can recompile haartraining.exe with OpenMP support, it will speed up process a little bit =).<\/p>\n Finally I got haarcascade.xml file. OpenCV perfomance.exe tool can be used for testing cascade.\u00a0 I tested it on my testing content\u00a0 – sorry for pun =).<\/p>\n\n\t\tSample\/Test data creation.<\/strong><\/h3>\n
\n
Training.<\/strong><\/h3>\n
Testing.<\/strong><\/h3>\n