notesum.ai

Published at November 26

cs.CV

cs.LG

Released Date: November 26, 2024

Authors: Ahmed Akl¹, Abdelwahed Khamis², Zhe Wang³, Ali Cheraghian², Sara Khalifa⁴, Kewen Wang³

Aff.: ¹Griffith University, Data61, CSIRO; ²Data61, CSIRO; ³Griffith University; ⁴Queensland University of Technology

Method	Backbones	VQA-CPv2 (%)				VQA-CP v1 (%)
Method		All	Y/N	Num	Other	All	Y/N	Num	Other
I Vanilla Backbones
SAN [48]	-	24.96	38.35	11.14	21.74	32.5	36.86	12.47	36.22
UpDn [3]	-	39.74	42.27	11.93	46.05	36.38	42.72	42.14	40.35
LXMERT [41]	-	48.66	47.49	22.24	56.52	52.82	54.08	25.05	62.72
II Answer Re-Ranking Methods
SCR [46]	UpDn	49.45	72.36	10.93	48.02	-	-	-	-
Loss-Rescaling [19]	LXMERT	66.4	79.77	59.06	61.41	69.47	86.82	45.03	62.03
III Ensemble Methods
RUBi [8]	UpDn	44.23	67.05	17.48	39.61	50.9	80.83	13.84	36.02
LMH [13]	UpDn	52.45	69.81	44.46	45.54	55.25	76.47	26.66	45.68
D-VQA [44]	LXMERT	69.75	80.43	58.57	69.75	-	-	-	-
GenB [11]	LXMERT	71.16	92.24	64.71	61.89	-	-	-	-
LBCL[24]	UpDn	60.74	88.28	45.77	50.14	61.57	84.48	42.84	46.32
BILI [51]	UpDn	59.76	88.2	41.76	49.79	62.59	85.29	46.77	46.37
FAN-VQA [7]	LXMERT	72.18	84.76	65.98	67.29	-	-	-	-
IV Data Augmentation Methods
CSS[9]	UpDn+LMH	58.95	84.37	49.42	48.21	60.95	85.6	40.57	47.03
Mutant [16]	UpDn	61.72	88.9	49.68	50.78	-	-	-	-
DGG [45]	UpDn	61.14	88.77	49.33	49.9	-	-	-	-
SIMPLEAUG [23]	LXMERT	62.24	69.72	53.63	60.69	-	-	-	-
SCLSM [47]	LXMERT	70.27	82.35	58.97	67.03	-	-	-	-
V Curriculum Learning Method (proposed)
$\text{TPCL}_{\text{Fix}\uparrow}$	LXMERT	75.83	91.55	68.49	69.61	76.78	90.74	72.22	64.72
$\text{TPCL}_{\text{Dyn}\uparrow}$	LXMERT	77.23	93.10	72.00	70.34	76.15	93.93	62.62	63.91