notesum.ai

Published at December 10

Test-time Correction with Human Feedback: An Online 3D Detection System via Visual Prompting

cs.CV

Released Date: December 10, 2024

Authors: Zetong Yang¹, Hanxue Zhang², Yanan Sun¹, Li Chen¹, Fei Xia³, Fatma Guney⁴, Hongyang Li¹

Aff.: ¹Shanghai AI Lab; ²Shanghai Jiao Tong University; ³Meituan Inc.; ⁴Koc University

Arxiv: http://arxiv.org/pdf/2412.07768v1

[Uncaptioned image]

Method	Backbone	Type	mAP (%) $\uparrow$	EDS (%) $\uparrow$
MonoDETR [67]	R101	Monocular	37.9	38.2
TTC-MonoDETR	R101	Monocular	42.6 (+5.7)	43.2 (+5.0)
MV2D [57]	R50	Multiview	38.9	38.3
TTC-MV2D	R50	Multiview	51.0 (+12.1)	51.0 (+12.7)
Sparse4Dv2 [31]	R50	Multiview + Temporal	38.8	37.6
TTC-Sparse4Dv2	R50	Multiview + Temporal	53.5 (+14.7)	52.1 (+14.5)
BEVFormer [27]	R101	BEV + Temporal	36.5	35.8
TTC-BEVFormer	R101	BEV + Temporal	47.8 (+11.3)	47.2 (+11.4)
BEVFormerV2-t8 [62]	R50	BEV +Temporal	39.6	38.9
TTC-BEVFormerV2-t8	R50	BEV +Temporal	51.6 (+12.0)	51.0 (+12.1)
RayDN [32]	R50	BEV + Temporal	39.6	38.7
TTC-RayDN	R50	BEV + Temporal	51.7 (+12.1)	50.8 (+12.1)
StreamPETR [54]	V2-99	BEV + Temporal	39.7	39.1
TTC-StreamPETR	V2-99	BEV + Temporal	52.3 (+12.6)	51.5 (+12.4)