notesum.ai

Published at November 27

EventCrab: Harnessing Frame and Point Synergy for Event-based Action Recognition and Beyond

cs.CV

Released Date: November 27, 2024

Authors: Meiqi Cao¹, Xiangbo Shu¹, Jiachao Zhang², Rui Yan³, Zechao Li¹, Jinhui Tang¹

Aff.: ¹Nanjing University of Science and Technology; ²Nanjing Institute of Technology; ³Nanjing University

Arxiv: http://arxiv.org/abs/2411.18328v1

Refer to caption

Dataset	Method	Accuracy (%)
Dataset	Method	Top-1	Top-5
PAF	HMAX SNN [52]	55.00	-
	STCA [13]	71.20	-
	Motion SNN [25]	78.10	-
	MST [50]	88.21	-
	Swin-T(BN) [50]	90.14	-
	EV-ACT [7]	92.60	-
	ExACT [58]	94.83	98.28
	Ours	96.49(+1.66)	100.00(+1.72)
SeAct	EventTransAct [2]	57.81	64.22
	EvT [38]	61.30	67.81
	ExACT [58]	67.24	75.00
	Ours	72.41(+5.17)	89.65(+14.65)
HARDVS	X3D	45.82	52.33
	SlowFast [4]	46.54	54.76
	ACTION-Net [49]	46.85	56.19
	R2Plus1D [45]	49.06	56.43
	ResNet18 [14]	49.20	56.09
	TAM [27]	50.41	57.99
	C3D [44]	50.52	56.14
	ESTF [48]	51.22	57.53
	Video-SwinTrans [28]	51.91	59.11
	TSM [24]	52.63	60.56
	TSCFormer [46]	53.04	62.67
	ExACT [58]	90.10	96.69
	Ours	97.11(+7.01)	99.48(+2.79)
DVS128 Gesture	Time-surfaces [30]	90.62	-
	SNN eRBP [18]	92.70	-
	Slayer [40]	93.64	-
	DECOLLE [19]	95.54	-
	EvT [38]	96.20	-
	TBR [16]	97.73	-
	EventTransAct [2]	97.92	-
	ExACT [58]	98.86	98.86
	Ours	98.80(-0.08)	100.00(+1.14)