Monocular 3D mesh prediction for cars in autonomous driving scenarios
CovenanteeUniversity of Toronto. Vector Institute
Document typeBachelor thesis
Rights accessOpen Access
In the context of autonomous driving cars, a necessity for precise 3D reconstructions of the road and environment is every day more explicit. The present work aims to expose an approach for reconstructing vehicles from a single (RGB) image of a road-like scenario. It is based on previous 3D mesh prediction algorithms that mix Convolutional Neural Network architectures for feature extraction with Graph Convolutional Networks. We build our own dataset and train our model to predict 3D meshes, providing the network with a single RGB image of a car and a 3D bounding box. We reveal how this method recovers precisely most of the geometric details of a car as well as shows successful levels of accuracy.