Hand gesture recognitions using vision based devices are user friendly but have self-occlusion problems which some fingers or parts of the hand are hidden from a device. This paper proposes estimation method of hand posture for self-occlusion when using a vision based device. A hidden finger is estimated by using Neural Network learned some finger constraints during hand motions. Some experiments show that estimation results are based on some finger constraints when the occlusion are continued for middle finger or ring finger.