Dataset generator for face detector training

identification, head pose, anthropometry

Many classification algorithms depend on a training dataset to be able to find appropriate parameters. The training dataset can significantly affect the resulting quality of the classifier hence it is necessary to use as good dataset for given task as possible. Either there already is an available ready-to-use dataset or a new one must be created. It is rather tedious and time consuming process to create a~new dataset from scratch because many algorithms usually require a large amount of tagged data. For many specific tasks the only way is to create a new dataset simply because there are no sets available. For the purposes of training and testing face detection algorithms we created a generator producing large number of images containing human head in various poses and positions placed in a predefined background. We are using such datasets to test success rate of head detection algorithms in dependence on the rotation of the head. Based on the test results we designed a tool to classify suitability of a security camera placement based on its parameters and space it should watch.