A deep learning method employs a neural network having three sub-nets to classify and retrieve the most similar 3D model of an object, given a rough 3D model or scanned images. The most similar 3D model is present in a database and can be retrieved to use directly or as a reference to redesign the 3D model. The three sub-nets of the neural network include one dealing with object images and the other two dealing with voxel representations. Majority vote is used instead of view pooling to classify the object. A feature map and a list of top N most similar well-designed 3D models are also provided.