DETopK: training for DETR based models with Top K predictions noise
Nhan Duy Quan, Dang Nguyen Chau · 2023
DETR is one of the first object detection models to apply the attention mechanism, which was proven to be very strong for many machine learning models. There have been many variants of DETR that can achieve competitive results on the COCO dataset, proving the DETR line’s capabilities of creating state-of-the-art models. In this paper, we present DETopK - a simple training method for DETR based models that focus on the detection area of object queries. Because of hardware limitation (2080Ti GPU), our experiment was performed with Deformable DETR multi-scale, on COCO-minitrain - a subset of the COCO dataset for a reasonable training time. It is very encouraging that the experimental results show that the proposed method helps Deformable DETR gain 0.27% in AP, 0.36% in AR and better accuracy overall when trained with our method.