Accurate Quaternion Polar Harmonic Transform for Color Image Analysis

Table 2

Comparison of retrieval performance in terms of mean average precision (mAP).

Method		Feature vector length	Oxford5K	Holidays	UK-bench
Hand craft-based methods	LCS + RN [22]		0.517	0.658	0.324
	Improved BOW [23]		0.546	0.747	0.352
	LTF [24]		—	0.689	0.355
	LBP + color [25]		0.664	0.735	0.347
CNN-based methods	Neural codes [26]	128	0.557	0.789	0.356
	CRB-M [27]	512	0.582	0.809	0.340
	VGG + GeM [28]	512	0.820	0.831	—
	AWF [29]	512	0.811	0.826	0.360
	Triplet network [30]	512	0.820	0.842	0.362
QPHT [31]		95	0.696	0.812	0.356
Our AQPHT		95	0.742	0.831	0.363

The bold values represent results of our proposed method, which denote the best results among the methods in Table 2.