A benchmark of CNN backbones on DINO-DETR performance in object detection

Abstract

Recent developments in DETR-based models have made significant improvements in training convergence but not small object detection. This paper combines the ConvNeXt and FocalNet backbones with DINO-DETR using timm and detrex, and presents a benchmark and analysis of the resulting model performances on MS-COCO and SODA-D. The results affirm many conclusions from the ConvNeXt and FocalNet papers while exhibiting inconsistencies for FocalNets on SODA-D. Finally, the results show encouraging performance for DINO-DETR with recent backbones on general object detection and the need for further improvement on small object detection with DINO-DETR across all backbones. Further efforts should be made to integrate state-of-the-art features from concurrent developments to produce new benchmarks on small object detection datasets with accessible existing technology.