notesum.ai

Published at November 29

FlowCLAS: Enhancing Normalizing Flow Via Contrastive Learning For Anomaly Segmentation

cs.CV

cs.LG

Released Date: November 29, 2024

Authors: Chang Won Lee¹, Selina Leveugle¹, Svetlana Stolpner², Chris Langley², Paul Grouchy², Jonathan Kelly¹, Steven L. Waslander¹

Aff.: ¹University of Toronto; ²MDA Space

Arxiv: http://arxiv.org/pdf/2411.19888v1

Refer to caption

Method	Inlier Labels	FS-L&F		Road Anomaly
Method	Inlier Labels	AUPRC $\uparrow$	$\text{FPR}_{95}$ $\downarrow$	AUPRC $\uparrow$	$\text{FPR}_{95}$ $\downarrow$
SynBoost [19]	✓	60.6	31.0	41.8	59.7
DenseHybrid [22]	✓	69.8	5.1	-	-
PEBAL [48]	✓	58.8	4.8	45.1	44.6
RbA [40]	✓	70.8	6.3	85.4	6.9
RPL [38]	✓	70.6	2.5	71.6	17.7
EAM [23]	✓	81.5	4.2	69.4	7.7
RWPM [54]	✓	71.2	6.1	87.3	5.3
UNO [16]	✓	81.8	1.3	88.5	7.4
FastFlow+DINOv2-L [52]	✗	8.9	94.4	57.1	91.8
FlowCLAS (ours)	✗	72.3	1.6	91.8	5.7