Detection of self intersection in synthetic hand pose generators

Synthetic hand pose data has been frequently used in vision based hand gesture recognition. However existing synthetic hand pose generators are not able to detect intersection between various hand parts and can synthesize self intersecting poses. Using such data may lead to learning wrong models. We propose a method to eliminate self intersecting synthetic hand poses by accurately detecting intersections between various hand parts. We model each hand part as a convex hull and calculate pairwise distance between the parts, labeling any pair with a negative distance as intersecting. A hand pose with at least one pair of intersecting parts is labeled as self intersecting. We show experimentally that our method is very accurate and performs better than existing techniques. We also show that it is fast enough for offline data generation.