Automated lip reading from videos requires lip segmentation. Threshold-based segmentation is straightforward, but it is rarely used. This study proposes a histogram threshold based on the feedback of shape information. Both good and bad lip segmentation examples were used to train an ϵ-support vector regression model to infer the segmentation accuracy from the region shape. The histogram threshold was optimised to minimise the segmentation error. The proposed method was tested on 895 images from 112 subjects using the AR Face Database. The proposed method, implemented in simple segmentation algorithms, reduced segmentation errors by 23.1%.