Machine Vision Training With Mixed Annotation Depth

The annotation process is segmented into two distinct types: first annotations providing detailed information (object masks, boundaries, positions) and second annotations providing less detailed information (object counts, presence/absence). This segmentation allows different annotation depths for different images in the training set, reducing overall annotation time while maintaining recognition accuracy through the combination of both annotation types during algorithm training.

Different images in the training set are assigned different annotation qualities locally. Some images receive comprehensive first annotations with detailed object information, while other images receive simplified second annotations with less detailed information. The image content recognition algorithm is trained to effectively utilize both annotation types, achieving reliable training results without requiring all images to have high-quality detailed annotations, thus improving annotation efficiency.