Training MaskDINO for instance segmentation
MaskDINO gives detection and instance masks from one transformer head, and on our defect data it beat a Mask R-CNN baseline once the dataset was clean. The config bits that actually mattered:
# start from the released COCO checkpoint, don't train the backbone from scratch
python train_net.py --num-gpus 1 `
--config-file configs/maskdino_swinl.yaml `
SOLVER.IMS_PER_BATCH 4 SOLVER.BASE_LR 1e-4 `
SOLVER.MAX_ITER 40000 MODEL.WEIGHTS maskdino_swinl_coco.pth