3-D Precision Positioning Based on Deep Comparison Convolutional Neural Networks

The UniShot 3D precision positioning model was developed using a deep comparison neural network (DCN). This dual-pipeline network extracts features from both the base and inquiry images in real time and predicts the observer's kinematic movements through internal comparison. We trained the model for transversal and depth movement detections and reported the precision and recall rates through static and dynamic experiments. We also analyzed the feature maps in the convolutional layers at various depths of the model to understand the comparison mechanism of the network. Results showed that the saliency feature patterns of DCNs are distinct from those of image recognition models and that the patterns for the transversal model were distinct from those for the depth model.