notesum.ai

Published at October 20

YOLO-RD: Introducing Relevant and Compact Explicit Knowledge to YOLO by Retriever-Dictionary

math.AG

cs.CC

Primary: 14N07, Secondary: 15A69, 14C05, 68Q17

Released Date: October 20, 2024

Authors: Hao-Tang Tsui¹, Chien-Yao Wang¹, Hong-Yuan Mark Liao¹

Aff.: ¹Institute of Information Science, Academia Sinica

Arxiv: https://arxiv.org/abs/2410.15346v1

Refer to caption

BackBone	Initializer	Params	$\|D\|$	$f$	$\text{mAP}^{val}_{.5:.95}$ (%)	$\text{mAP}^{val}_{.5}$ (%)
YOLOv7	Origin	37.2M	-	-	50.04	68.65
YOLOv9	Origin	25.3M	-	-	52.64	69.56
Faster-RCNN	Origin	43.1M	-	-	38.40	59.00
Deformable DETR	Origin	40.1M	-	-	43.80	62.60
YOLOv7	VM	37.4M	512	$\mathbb{R}^{512}$	51.34 ( $\uparrow$ 2.61%)	69.37 ( $\uparrow$ 1.05%)
YOLOv9	VM	25.8M	512	$\mathbb{R}^{512}$	53.41 ( $\uparrow$ 1.46%)	70.57 ( $\uparrow$ 1.46%)
Faster-RCNN	VM	44.1M	512	$\mathbb{R}^{512}$	40.50 ( $\uparrow$ 5.47%)	60.30 ( $\uparrow$ 2.20%)
Deformable DETR	VM	41.2M	512	$\mathbb{R}^{512}$	44.10 ( $\uparrow$ 0.68%)	63.30 ( $\uparrow$ 1.12%)
YOLOv7	VLM	37.4M	512	$\mathbb{R}^{512}$	51.72 ( $\uparrow$ 3.37%)	70.12 ( $\uparrow$ 2.15%)
YOLOv9	VLM	25.8M	512	$\mathbb{R}^{512}$	53.36 ( $\uparrow$ 1.37%)	70.55 ( $\uparrow$ 1.43%)
Faster-RCNN	VLM	44.1M	512	$\mathbb{R}^{512}$	40.50 ( $\uparrow$ 5.47%)	60.40 ( $\uparrow$ 2.37%)
Deformable DETR	VLM	41.2M	512	$\mathbb{R}^{512}$	44.40 ( $\uparrow$ 1.37%)	63.30 ( $\uparrow$ 1.12%)
YOLOv7	LLM	38.2M	512	$\mathbb{R}^{1024}$	51.33 ( $\uparrow$ 2.59%)	69.40 ( $\uparrow$ 1.10%)
YOLOv9	LLM	25.8M	512	$\mathbb{R}^{1024}$	53.28 ( $\uparrow$ 1.22%)	70.48 ( $\uparrow$ 1.33%)
Faster-RCNN	LLM	44.6M	512	$\mathbb{R}^{1024}$	40.70 ( $\uparrow$ 5.99%)	60.80 ( $\uparrow$ 3.05%)
Deformable DETR	LLM	41.7M	512	$\mathbb{R}^{1024}$	44.20 ( $\uparrow$ 0.91%)	63.10 ( $\uparrow$ 1.12%)