Abstract: In example embodiments, a software application may implement techniques for detection and identification of physical objects such as signs in images. The techniques may a) decouple detection of the physical objects (e.g., signs) and their meta-classes from identification of information about the physical objects such as their specific technical name, b) use weakly supervised training to train a ML detection model to detect the physical objects (e.g., signs) and their meta-classes using image-wise meta-class labeled training images and/or c) use a combination of textual clues from OCR and visual clues from keypoint matching of image descriptors to identify the information (e.g., specific technical name) of each physical objects (e.g., sign) once it has been assigned a meta-class.