基於深度學習之視覺定位-以故宮南院為例

Visual localization uses images to regress camera position and orientation. It has many applications in computer vision such as autonomous driving, augmented reality (AR) and virtual reality (VR), and so on. The convolutional neural network simulates biological vision and has a good image feature extraction ability, so using it in visual localization can improve regression accuracy. Although our team has built an image indoor localization model for Southern Branch of the National Palace Museum, this paper tries to use new network and loss function to achieve better positioning accuracy. In this paper, we use ResNet-50 as backbone network, and change the output layer to 3-dimensional position and 4-dimensional orientation quaternion, and use learnable weights loss function. We compare different pretrained models and normalization methods, and the best result improves the positioning accuracy by about 60%.