Systems, methods, and computer readable media to improve the operation of detecting contact between a finger or other object and a surface are described. In general, techniques disclosed herein utilize a depth map to identify an object and a surface, and a classifier to determine when the object is touching the surface. Unlike the prior art, a measure of the object's “distance” is made relative to the surface and not the camera(s) thereby providing some measure of invariance with respect to camera pose. The object-surface distance measure can be used to construct an identifier or “feature vector” that, when applied to a classifier, generates an output indicative of whether the object is touching the surface. The classifier may be based on machine learning and can be trained off-line before run-time operations are commenced. In some embodiments, temporal filtering may be used to improve surface detection operations.