Survey on deep learning with class imbalance

Journal of Big Data

Table 18 Summary of data sets and class imbalance levels

Paper	Data sets	Data type	Class count	Data set size	Min class size	Max class size	\(\rho\) (Eq. 1)
[79]	CIFAR-10	Image	10	60,000	2340	3900	2.3
[20]	WHOI-Plankton	Image	103	3,400,000	< 3500	2,300,000	657
[21]	Public cameras	Image	19	10,000	14	6986	499
[18]	CIFAR-100 (1)	Image	2	6000	150	3000	20
	CIFAR-100 (2)	Image	2	1200	30	600	20
	CIFAR-100 (3)	Image	2	1200	30	600	20
	20 News Group (1)	Text	2	1200	30	600	20
	20 News Group (2)	Text	2	1200	30	600	20
[88]	COCO	Image	2	115,000	10	100,000	10,000
[103]	Building changes	Image	6	203,358	222	200,000	900
[89]	GHW	Structured	2	2565	406	2159	5.3
[89]	ORP	Structured	2	700	124	576	4.6
[19]	MNIST	Image	10	70,000	600	6000	10
	CIFAR-100	Image	100	60,000	60	600	10
	CALTECH-101	Image	102	9144	15	30	2
	MIT-67	Image	67	6700	10	100	10
	DIL	Image	10	1300	24	331	13
	MLC	Image	9	400,000	2600	196,900	76
[90]	KEEL	Structured	2	3339	26	3313	128
[91]	CIFAR-10	Image	10	60,000	250	5000	20
[91]	CIFAR-100	Image	100	60,000	25	500	20
[22]	CelebA	Image	2	160,000	3200	156,800	49
[117]	MNIST	Image	10	60,000	50	5000	100
	MNIST-back-rot	Image	10	62,000	12	1200	100
	CIFAR-10	Image	10	60,000	5000	5000	1
	SVHN	Image	10	99,000	73	7300	100
	STL-10	Image	10	13,000	500	500	1
[118]	CelebA	Image	2	160,000	3200	156,800	49
[92]	EmotioNet	Image	2	450,000	45	449,955	10,000
[23]	MNIST	Image	10	60,000	1	5000	5000
	CIFAR-10	Image	10	60,000	100	5000	50
	ImageNet	Image	1000	1,050,000	10	1000	100

Images from CelebA and EmotioNet are treated as a set of binary classification problems, because they are each annotated with 40 and 11 binary attributes, respectively. The COCO data class imbalance arises from the extreme imbalance between background and foreground concepts