This paper presents a novel feedforward neural network for sensor-based dynamic hand gesture recognition. The algorithm, termed PairNet, is capable of carrying out accurate gesture spotting for the sensory data produced by basic accelerators and gyroscopes, which are commonly deployed in internet of things devices. The gesture classification outcomes are then obtained from the spotting results by the Maximum A Posteriori (MAP) estimation. To illustrate the effectiveness of the proposed algorithm, a prototype system based on a mobile phone has been implemented. Experimental results reveal that, while attaining realtime operations, the proposed algorithm has superior accuracy over existing sensor-based counterparts for hand gesture recognition.