Object Viewpoint Estimation using CNN-based Classifier

The pose information of an object plays an important role in the field of robotics and computer vision. The pose estimation process is conventionally difficult and complicated. In the case of pose estimation that requires 3D CAD models, it is not easy to obtain the CAD models for target objects in specific tasks. In this paper, we show that the viewpoint can be estimated through an image classification method instead of complex pose estimation approaches. We reorganize a dataset to predict an azimuth angle based on Pascal 3D+ dataset. Several convolution neural networks-based classifiers are applied for categories discretized with viewpoint. The performance of EfficientNet is the best, and the smaller the number of discretized viewpoint bins, the better the performance.