In HALCON 12, the sample-based identification (sbi) tool has been improved to accurately identify objects despite cluttered backgrounds or strong local illumination variations. The objects shown on the right are not accurately identified with the old version of sbi, however HALCON 12’s improved sbi tool is successful in properly identifying them. It is also now possible to train multiple objects using any number of images for each class. In the HALCON 11 version it was necessary to use approximately the same number of training images for each object to acheive the best results. This is no longer necessary.
MVTec tested the improved sample-based identification tool on a large, complex test set which contained 100 different objects and 25,000 images with challenges such as scaling changes, local illumination variations, occlusions, cluttered backgrounds, de-focusing, deformations, etc. Using a total of 1,600 training images, the HALCON 12 sample-based identification tool performed with a recognition rate of 90.4%. Using the old HALCON 11 sample-based identification tool, the recognition rate on the same test set was 86.8%.
To see the improved sample-based identification tool in action, watch Managing Director, Olaf Munkelt demonstrate and discuss the technology here.