As deep learning methods won the huge success in 2D area, they're striding forward into 3D computer vision community during the last few years, meanwhile as point cloud data becoming ubiquitous with rapid development of 3D sensors, deep learning implemented on 3D point cloud steps into different scenarios as 3D classification, segmentation, object detection and reconstruction. However, processing 3D point cloud data by deep learning is a non-trivial job because of its naturally irregular data format. Some predecessors overcame the key issue by designing a direct deep network, left the problem neglecting implicit local feature relations as an open question, though. Our proposal network, dual-channel multi-layer perceptron deep network, or DualNet, aiming to take advantage of local features and to learn from combinatorial global and local features, thus to achieve both stable order invariance and high accuracy for 3D object classification tasks. A couple of experiments will be conducted to verify the performance of our proposal DualNet by comparing with state-of-the-art methods on major 3D point cloud database.