Mouth map-assisted localisation of ASM lip landmarks

摘要

The extended active shape models have shown their robustness in facial landmarks fitting. However, we often meet substantial errors of lip landmarks fitting mostly due to teeth boundaries. An attempt to reduce the errors using high-quality mouth maps has been made recently. Since most existing facial datasets cannot produce high-quality mouth maps, use of mouth maps should be careful not to bring even more serious errors from them. Unlike the existing scheme using the mouth map, we use the mouth map not for profile model but only for better initialisation of landmarks right before profile matching. As a result, the proposed method attains higher accuracies compared to the original Stasm 4 and the existing method using the mouth map. Experimental studies validate the effectiveness of the proposed method. Since the proposed method uses the mouth map only for initialisation, it requires neither additional task in training nor high-quality mouth maps.