One of the main challenges in 'real-life' object recognition applications is keeping some invariance properties such as: translation, scaling, and rotation. However, trying to maintain such invariants can impair recognition capabilities, especially when the family of objects of interest has a large shape variability. We present a general family of shape metrics that generalizes Procrustes metric and within this framework learns the desired shape metric parameters from labeled training samples. The learnt distance retains invariance properties on one hand and emphasizes the discriminative shape features on the other hand. We show how these metrics can be incorporated in multi-class classification kernel SVMs. We demonstrate the merits of this approach on multi-class facial expressions recognition using the AR dataset. The results address some questions and cautions regarding the interpretation of classification results when using still images datasets collected in a controlled lab environment and their relevance for 'real-life' applications.