DETA is a new model that improves upon Deformable DETR. Equipped with a Swin-large backbone and pre-trained on Object 365, it obtains an average precision (AP) of 63 on COCO.

DetaForObjectDetection has the same API as models like DETR, Deformable DETR and Conditional DETR, hence I refer to my notebooks here regarding fine-tuning on custom data.

This folder just contains a notebook illustrating inference.