At a glance
STK is a software tool for training and testing HOG/SVM-based classifiers. It was designed to support users in complex operations, such as the learning process for an automated classifier, in a simple and effective way. With STK anyone can undertake the task of training a classifier: not only computer vision experts. STK natively supports Detecto IP as actual target.
- Media annotation. In a STK project you can use videos (supported formats: .dv, .avi, .mov, .wmv, .mpeg2, .mp4) or collections of images (supported formats: .jpg, .jpeg, .bmp, .png, .tiff, .ppm, .pgm) for classification, training and testing. To annotate media, STK supports the concept of layers as collection of boxes referred to a specific object of interest (e.g. pedestrian, motorcycle, car, traffic sign, animal, cell, etc.). To speed up the manual annotation, STK allows you to select boxes (key box nodes) in a few frames and then automatically generate the boxes in the frames in between. This operation is done by linear interpolation on the geometry of the key box nodes. A project can contain more than one layer, for example: one for ground truth annotations and one for the detection results from a trained classifier. STK provides facilities to ease layers handling such as their merging ('simple join' mode / 'object grouping' mode).
- Training. STK exposes a guided procedure through all the training phases. The procedure relies on a training set of annotated media for positive and negative images and/or binary datasets (features previously computed and exported form STK after training process). The guided procedure embraces fine tuning for training parameters and can display a qualitative representation of the computed classifier that will be able to detect objects of interest in project video/images.
- ROI (Region of Interest). In a STK project you can define a region on the video frames/images in which objects of interest are supposed to be. Different scales of a ROI are usually used for each video frame/image to let the classifier detecting objects at different scales (image pyramid). STK supports ROIs that apply to different media and can be used with different classifiers (on condition the template size is the same). ROIs can be created manually, specifying minimum/maximum scale and density, or automatically starting from camera parameters.
- Detection. STK can exercise the detection process running a classifier on a sliding window over the input image. In order to detect objects at various sizes and ranges, the detection process is repeated on successively scaled copies of the image. If you have any constraints on the scene you can restrict the search area setting a ROI.
- Performance analysis. STK let you evaluating the classification results of a classifier. Provided at least one layer with ground truth data and one layer with detection results (and optionally a layer containing regions to be ignored), the tool can compute standard metrics about performances for a classifier such as precision, recall, F1-score, TP, FP, FN.
A 15 days trial version for Windows is available free of charge. We recommend those who are using the software for the first time to download the tutorial archive as well. This tutorial and related test projects will run you through the whole process of training a classifier for no entry signs with STK.
The development of STK has been sponsored by Xilinx, Inc.
For more information about STK please contact
To request a quote or place an order please contact