notesum.ai

Published at November 22

Harlequin: Color-driven Generation of Synthetic Data for Referring Expression Comprehension

cs.CV

cs.CL

Released Date: November 22, 2024

Authors: Luca Parolari¹, Elena Izzo¹, Lamberto Ballan¹

Aff.: ¹University of Padova, Italy

Arxiv: http://arxiv.org/abs/2411.14807v1

Refer to caption

TransVG [7]:
Method	RefCOCO			RefCOCO+			RefCOCOg
Method	val	testA	testB	val	testA	testB	val	test
Real	63.33	69.05	55.62	64.69	69.02	55.76	64.04	63.22
Synth $\rightarrow$ Real	65.77	70.66	56.80	66.66	72.01	55.66	65.13	64.33
(Improv.)	+2.44	+1.61	+1.18	+1.97	+2.99	-0.10	+1.09	+1.11
VLTVG [36]:
Real	69.66	74.33	61.35	70.83	76.02	61.71	70.57	70.03
Synth $\rightarrow$ Real	69.60	75.76	61.14	71.46	77.16	61.30	70.04	69.57
(Improv.)	-0.06	+1.43	-0.21	+0.63	+1.12	-0.41	-0.53	-0.46
LGR-NET [25]:
Real	82.71	85.77	79.31	71.11	75.45	63.35	70.75	71.11
Synth $\rightarrow$ Real	84.38	87.13	80.67	71.40	75.60	64.70	74.61	75.22
(Improv.)	+1.67	+1.36	+1.36	+0.29	+0.15	+1.35	+3.86	+4.11