5. Conclusion and outlook

Shape recognition techniques, which in the 1980s were based on simple geometric characterization criteria, have since the 2000s been supported by point-of-interest detection and descriptor calculation techniques. The underlying algorithms are fast and parallelizable, making them suitable for use in embedded systems, even those with very low computing capacity (and also very low power consumption).

Since around 2020, so-called deep learning techniques have enabled semantic segmentation (detection of objects of the same type) or instance-based segmentation (detection of each object independently of the others), bringing real meaning to points of interest that are not necessarily linked to a visually perceived object. By adding this segmentation to the descriptors, the applications presented will probably be considered solved problems in less than ten years' time,...