CPC G06V 10/7788 (2022.01) [G06N 3/09 (2023.01); G06N 3/091 (2023.01); G06V 10/267 (2022.01); G06V 10/763 (2022.01); G06V 10/774 (2022.01); G06V 20/70 (2022.01)] | 20 Claims |
1. A computer-implemented method of adaptive learning for semantic segmentation of images, the method comprising:
obtaining a semantic segmentation model, the semantic segmentation model having been trained using a first subset of sliding windows and corresponding first ground truth masks for the first subset of sliding windows;
ranking a plurality of sliding windows from a corpus of training images according to an uncertainty metric;
selecting a next subset of sliding windows from the corpus of training images based on the ranking and based on a similarity metric for one or more characteristics of a sliding window relative to other sliding windows;
providing a collaborative user interface for labeling the next subset of sliding windows;
receiving ground truth masks for the next subset of sliding windows using the collaborative user interface; and
retraining the semantic segmentation model using the next subset of sliding windows and the ground truth masks for the next subset of sliding windows.
|