Have you tried out how DETReg matches up against using a checkpoint pre-trained on e.g. MS COCO on small to medium sized datasets? The way I understand the supervised baseline is that only the backbone is pre-trained on a classification task.
Specifically, I would be interested in seeing if DETReg or DETReg + MS COCO fine-tuning can outperform transfer learning based on MS COCO.
Hi, thanks for the great work!
Have you tried out how DETReg matches up against using a checkpoint pre-trained on e.g. MS COCO on small to medium sized datasets? The way I understand the supervised baseline is that only the backbone is pre-trained on a classification task.
Specifically, I would be interested in seeing if DETReg or DETReg + MS COCO fine-tuning can outperform transfer learning based on MS COCO.