Context-Aware and Locality-Constrained Coding for Image Categorization

Table 2

The comparison result with several coding styles on Caltech101 (training examples with size of 30), Caltech256 (training examples with sizes of 30, 60), and Scene15 (training examples with size of 100). Up the bold line are the results from the corresponding literature; below the line are the results from our implementation. And the two versions of CALC implemented are dictionary trained using 𝑘-means and Algorithm 2.

Unit: %	Cal. 101 (# 30)	Cal. 256 (# 30)	Cal. 256 (# 60)	Scene15 (# 100)
VQ [19]	64.60 ± 0.80	NA	NA	81.40 ± 0.50
SC [11]	73.20 ± 0.54	34.02 ± 0.35	40.14 ± 0.91	80.28 ± 0.93
LSC [16]	NA	35.74 ± 0.10	40.32 ± 0.32	89.78 ± 0.40
LLC [15]	73.44 ± NA	41.19 ± NA	47.68 ± NA	NA
LSVQ [14]	74.21 ± 0.81	NA	NA	82.70 ± 0.39
LCSR [17]	73.23 ± 0.81	NA	NA	87.23 ± 1.14
LLC [ours]	72.32 ± 0.91	40.32 ± 0.26	46.56 ± 0.78	81.73 ± 0.75
LSVQ [ours]	72.58 ± 0.79	38.51 ± 0.42	43.10 ± 0.11	83.08 ± 0.56
CALC (𝑘-means)	74.90 ± 0.44	42.37 ± 0.38	49.45 ± 0.67	81.89 ± 0.54
CALC (learned)	75.84 ± 0.56	43.12 ± 0.62	51.44 ± 0.92	82.53 ± 0.81