notesum.ai

Published at November 26

DOGE: Towards Versatile Visual Document Grounding and Referring

cs.CV

cs.AI

Released Date: November 26, 2024

Authors: Yinan Zhou¹, Yuxin Chen², Haokun Lin³, Shuyu Yang¹, Li Zhu¹, Zhongang Qi², Chen Ma⁴, Ying Shan²

Aff.: ¹Xi'an Jiaotong University; ²ARC Lab, Tencent PCG; ³Institute of Automation, CAS; ⁴City University of Hongkong

Arxiv: http://arxiv.org/abs/2411.17125v1

[Uncaptioned image]

	Grounding						Grounding-and-Referring						Referring
Document Type	${G_{a}}$		${G_{r}}$		${G_{o}}$		${GR_{a}}$		${GR_{r}}$		${GR_{o}}$		${R_{t}}$
	${Acc}$	${F1_{all}}$	${Acc}$	${F1_{all}}$	BLEU4	${F1_{all}}$	${Acc}$	${F1_{all}}$	${Acc}$	${F1_{all}}$	BLEU4	${F1_{all}}$	${Acc}$
Poster	82.5	77.5	72	65.7	43	41.6	87.5	85.8	74.5	56.6	45.6	25.8	72.5
Chart	91.5	74	70.5	43.9	32.5	46.1	79.3	60.3	65.9	57.6	39.5	54.4	61
PDF Document	80.3	71	65	45	35.9	29.1	72.6	50	67.3	58.1	37.3	14.7	19.1
Average	84.8	74.2	69.2	51.5	37.1	41.9	79.8	65.4	69.2	57.4	40.8	31.6	50.9