Logo Detection With No Priors

In recent years, top referred methods on object detection like R-CNN have implemented this task as a combination of proposal region generation and supervised classification the proposed bounding boxes. Although pipeline has achieved state-of-the-art results in multiple datasets, it inherent limitations that make very complex inefficient computational terms. Instead considering standard strategy, paper we enhance Detection Transformers (DETR) which tackles set-prediction problem directly an end-to-end fully differentiable without requiring priors. particular, incorporate Feature Pyramids (FP) to DETR architecture demonstrate effectiveness resulting DETR-FP approach improving logo thanks improved small logos. So, any domain specific prior be fed model, obtains competitive OpenLogo MS-COCO datasets offering relative improvement up 30%, when compared Faster baseline strongly depends hand-designed