Computational strategies for skin detection

In this paper we compare different computational strategies for skin detection. They differ in the type of data used in the training phase, the type of pre-processing done on the query image, and the level of visual information used. In particular, we define a high-level computational strategy, which uses a face detector in the pre-processing step. Two different implementations of it are proposed: one relies on an adaptive single gaussian model, the other a fixed threshold skin cluster detector on an illuminant-independent image representation. The experimental results on a heterogeneous dataset containing images acquired under uncontrolled lighting conditions show that the high-level strategies outperform low-level ones.