Training: There is no training step.

Onboarding: For each object, 43 random pre-rendered images from the "train_pbr" dataset are selected. The masks of these objects are then extracted and encoded using CLIP. This data is input into a simple expert binary classifier, trained for each object.

Test: During testing, Fastsam-s extracts masks from a test image, and the object's ID is determined by the expert binary classifiers.