The XAMI model integrates two key components: a detector (based on YOLO or RT-DETR models) and a segmentor which relies on the SAM architecture. For optimal performance, we train these components separately. This approach allows for dedicated training of the detector and the segmentor, followed by combined training where the detector's layers are frozen.
The individual_train.ipynb notebook provides step-by-step instructions on how to train these models separately using various configurations. These steps can be skipped if you plan to use the pre-trained checkpoints. The train_combined.ipynb notebook shows how to train the segmentor using detector-generated bounding boxes.