A Boosting-Based Deep Distance Metric Learning Method

Table 1

Recall@K (%) on CUB-200-2011.

	CUB-200-2011
R@	1	2	4	8	16	32
Contrastive [2]	26.4	37.7	49.8	62.3	76.4	85.3
Triplet [2]	36.1	48.6	59.3	70.0	80.2	88.4
LiftedStruct [2]	47.2	58.9	70.2	80.2	89.3	93.2
N-Pairs [11]	51.0	63.3	74.3	83.2	—	—
Angular loss [12]	54.7	66.3	76.0	83.9	—	—
HDC [28]	53.6	65.7	77.0	85.6	91.5	95.5
HTL [16]	57.1	68.8	78.7	86.5	92.5	95.5
ABE [9]	60.6	71.5	79.8	87.4	—	—
DREML [10]	63.9	75.0	83.1	89.7	—	—
MS [13]	65.7	77.0	86.3	91.2	95.0	97.3
SoftTriple [29]	65.4	76.4	84.5	90.4	—	—
Ours	69.6	80.3	87.2	93.5	97.6	98.8

Note. DHC and HTL are short for hybrid dilated convolution and hybrid transfer learning, respectively.