Conditional Random Fields for Image Labeling

Table 1

Quantitative comparison on Leuven dataset of [1]. The table compares the average time per image and performance (object and stereo labeling accuracy) of joint object and stereo algorithms, using graph cut + range-move (GC + Range ()), an extension of cost-volume filtering, and [1]’s dense CRF with higher-order terms and filter-based inference (with and without cost-volume filtered unary, and using different approaches). HO means higher-order terms of [1] in the table.

Algorithm	Time (s)	Object (% correct)	Stereo (% correct)
GC + Range () [4]	24.6	95.94	76.97
GC + Range () [4]	49.9	95.94	77.31
GC + Range () [4]	74.4	95.94	77.46
Extended CostVol ([39] filter)	4.2	95.20	77.18
Dense + HO ([39] filter)	3.1	95.24	78.89
Dense_HO ([58] filter)	2.1	95.06	78.21
Dense + HO + CostVol ([58] filter)	6.3	94.98	79.00