Assessing the 3D Position of a Car with a Single 2D Camera Using Siamese Networks

Using computer vision for the classification of an object’s 3D position using a 2D camera is a topic that has received some attention from researchers over the years. Visual data is interpreted by the computer to recognize the objects found. In addition, it is possible to infer their orientation, evaluating their spatial arrangement, rotation, or alignment in the scene. The work presented in this paper describes the training and selection of a siamese neural network for classifying the 3D orientation of cars using 2D images. The neural network is composed of an initial phase for feature selection through convolutional neural networks followed by a dense layer for embedding generation. For feature selection, four architectures were tested: VGG16, VGG19, ResNet18 and ResNet50. The best result of 95.8% accuracy was obtained with the VGG16 and input images preprocessed for background removal.