Mesh models have been widely employed in current CAD/CAM systems, where the workpiece is considered as made up of a number of features limited by natural boundaries. The natural boundaries among features are in most cases edges where an abrupt change of point differential geometry properties occurs. However, such features and underlying surface portions could also be connected smoothly without an abrupt change at the natural boundaries. In the context of ISO GPS (Geometrical Product Specifications and Verification) Standards, partitioning is a fundamental operation that decomposes a mechanical part into independent surface portions for a functional specification purpose. In this paper, an enhanced mesh partitioning method is proposed to enable a feature-based decomposition considering kinematic invariance classes. The proposed two-step method includes an initial partitioning based on sharp edge detection and an enhanced partitioning process based on non-sharp edge rectification. The partitioning criteria rely on two surface descriptors derived from shapes’ principal curvatures: Curvedness and Shape Index. To refine the boundary points on non-sharp edges, conformal geometry is exploited during the enhanced partitioning process by mapping the 3D surface onto a 2D unit disk. Connecting regions without abrupt changes at their natural boundaries are well partitioned after the boundary rectification process. A statistical evaluation process is used to address invariance class identification for each partitioned surface portion on the part. Experiments and results on different mesh models are presented to demonstrate the effectiveness of invariance class partitioning for ISO GPS. •A two-step enhanced partitioning process for ISO GPS is developed.•The interpretation of the feature principle in a systematic and algorithmic way is presented.•A robust evaluation of existing discrete curvature estimation methods is described•Two surface descriptors, Curvedness and Shape Index are adopted for partitioning.•Conformal geometry is used to solve the partitioning problem in a lower dimension