Nikhil Jangamreddy

Parametric Deep Neural Networks for Computer Vision Problems

There are multiple traditional geometric problems in computer vision such as camera pose estimation, 3D reconstruction depth estimation, and tracking for which strong mathematical models and efficient, robust algorithms have been known for decades. Recently with the popularity of deep learning, researchers have applied various deep neural networks models to almost all of these problems and demonstrated state of the art performance on the benchmark datasets. However, modern learning based techniques also have known shortcomings in terms of handling out of distribution situations and susceptibility to adversarial attacks. The objective of this project is to explore parametric deep neural networks, whereby a standard deep neural network is constrained to broadly follow the steps of a prescribed mathematical model with a scope for learning each of the individual step using the available data. The developed model should ideally have the flexibility of learning based techniques, and at the same time robustness and interpretability of conventional mathematical techniques.