notesum.ai

Published at December 3

Noisy Ostracods: A Fine-Grained, Imbalanced Real-World Dataset for Benchmarking Robust Machine Learning and Label Correction Methods

cs.LG

cs.CV

Released Date: December 3, 2024

Authors: Jiamian Hu¹, Yuanyuan Hong¹, Yihua Chen¹, He Wang², Moriaki Yasuhara¹

Aff.: ¹The University of Hong Kong; ²Nanjing Institute of Geology and Palaeontology, CAS

Arxiv: http://arxiv.org/pdf/2412.02313v1

Refer to caption

	ResNet-50				ViT-B-16
Method	Acc	P@	R@	F1@	Acc	P@	R@	F1@
CE	95.98	88.50	77.80	79.51	95.03	83.31	75.30	76.64
CE+mixup	95.11	74.96	69.83	69.42	90.33	57.99	51.95	52.98
CL	95.16	80.97	69.01	71.51	90.62	57.48	53.27	53.75
Loss-clip	95.79	81.69	72.78	74.09	93.71	72.14	64.27	65.26
Co-teaching	95.79	79.01	71.79	73.23	94.71	77.77	71.69	72.37
Co-teaching+	96.19	84.09	77.57	78.21	91.82	66.04	63.45	63.65
DivideMix	95.50	53.42	57.33	54.86	84.79	23.00	28.22	24.92
SAM	94.84	78.93	68.08	70.60	93.51	77.89	64.11	69.82
PLM	96.77	77.77	72.74	73.46	96.36	81.28	75.17	76.41
SDN	95.11	78.13	71.31	72.04	92.84	76.49	62.16	64.96
LW	95.50	86.59	74.28	76.93	94.65	77.15	69.61	71.35
Transition*	84.22	61.80	55.33	56.49	83.28	62.64	52.95	54.18