notesum.ai

Published at November 18

Leveraging MLLM Embeddings and Attribute Smoothing for Compositional Zero-Shot Learning

cs.CV

cs.AI

Released Date: November 18, 2024

Authors: Xudong Yan¹, Songhe Feng¹, Yang Zhang¹, Jian Yang², Yueguan Lin², Haojun Fei²

Aff.: ¹School of Computer Science and Technology, Beijing Jiaotong University; ²Qifu Technology

Arxiv: http://arxiv.org/abs/2411.12584v1

Refer to caption

		MIT-States					C-GQA					VAW-CZSL
Method		$AUC$	$HM$	$Seen$	$Unseen$		$AUC$	$HM$	$Seen$	$Unseen$		$AUC$	$HM$	$Seen$	$Unseen$
SymNet [15]		3.2	13.7	22.7	20.1		1.9	10.8	20.3	11.8		2.8	13.5	20.2	18.0
CompCos [19]		12.3	28.2	39.0	39.5		5.0	17.7	32.8	19.1		6.5	20.8	30.5	27.4
Co-CGE [20]		10.3	25.1	41.0	33.1		4.2	15.2	32.9	17.0		6.2	19.7	31.0	26.1
SCEN [13]		9.8	24.6	35.1	36.5		3.8	15.3	31.5	15.7		5.7	19.2	29.9	24.5
OADis [38]		13.1	29.0	42.3	27.3		2.3	12.1	23.3	12.8		4.1	16.2	26.0	20.7
INV [44]		11.5	26.6	28.5	25.0		1.4	7.9	28.6	6.8		2.0	11.1	21.1	11.9
CANet [42]		13.6	29.8	46.4	39.9		5.7	18.9	34.8	20.5		6.7	21.0	31.2	27.4
ProCC [10]		9.5	28.1	43.1	39.1		3.5	15.1	32.4	15.8		3.6	18.9	26.9	25.5
CLIP [28]		11.0	26.1	30.2	46.0		1.4	8.6	7.5	25.0		-	-	-	-
CoOp [28]		13.5	29.8	34.4	47.6		4.4	17.1	20.5	26.8		-	-	-	-
TRIDENT (Ours)		14.2	30.9	44.5	40.0		8.0	22.6	39.5	24.1		8.3	23.4	33.3	31.1