notesum.ai

Published at December 5

MVUDA: Unsupervised Domain Adaptation for Multi-view Pedestrian Detection

cs.CV

Released Date: December 5, 2024

Authors: Erik Brorsson¹, Lennart Svensson², Kristofer Bengtsson¹, Knut Åkesson

Aff.: ¹Global Trucks Operations, Volvo Group, Gothenburg, Sweden; ²Department of Electrical Engineering, Chalmers University of Technology, Gothenburg, Sweden

Arxiv: http://arxiv.org/pdf/2412.04117v1

Method	MODA	MODP	Precision	Recall
MultiviewX $\rightarrow$ Wildtrack
^†Lima et al. [lima2024mean]	85.1	74.8	93.9	91.0
^†PPM [qiu2024ppm]	90.3	72.6	94.4	96.0
Oracle	91.3	75.5	97.0	94.2
\hdashline GMVD [gmvd]		70.7	73.8	89.1	80.6
TMVD [deeptopdown]	74.9	76.9	90.4	83.8
MVFP [mvfp]	82.6	76.2	89.6	93.4
Baseline	70.0	73.6	89.2	79.6
MVUDA (ours)	85.4	75.3	96.5	88.7
Wildtrack $\rightarrow$ MultiviewX
^†Lima et al. [lima2024mean]	75.9	78.6	96.2	79.0
Oracle	91.2	82.1	97.5	93.6
\hdashline Baseline		35.9	66.4	82.8	45.2
MVUDA (ours)	82.4	75.4	93.3	88.8