EURASIP Journal on Advances in Signal Processing

Table 7 Patch-level performances of CRF learning by the likelihood maximization (ML) and margin maximization (MM) methods, using only SIFT and color descriptors on MSRC-9 class data.

From: Scene Segmentation with Low-Dimensional Semantic Representations and Conditional Random Fields

Method	ML/SP_LBP		MM/ICM		MM/MP_LBP		MM/TRWS		MM/GC		MM/FastPD
	EF-1	EF-2	EF-1	EF-2	EF-1	EF-2	EF-1	EF-2	EF-1	EF-2*	EF-1	EF-2
Accuracy (%)	76.3	78.9	75.4	77.9	76.8	79.3	77.0	79.5	76.9	–	76.9	79.6
CPU time (s)	6954.5	7288.4	307.4	205.4	5041.9	3290.2	2471.4	2122.9	140.8	–	100.2	150.9

*GC failed here because SVMStruct does not enforce hard constraints on the weights and requested a weight update that produced a non-submodular energy function [45]. This could be fixed, but, given the poor performance of GC, we did not pursue this.

Back to article page