A Comparative Study of MonoLoco with Improvised Loss Functions

Abstract: The native concept of 3D human localization, from monocular color images is an ill posed problem. Considering the limitations of neural networks, we compare various loss functions that are based on a variety of distributions. Monoloco is a 3D pedestrian localization architecture which uses a lightweight feed forward neural network and predicts the distance of pedestrians, from the camera and the uncertainty associated with its prediction based on Laplacian loss. We trained two individual models on Kitti dataset with updated unnormalization methods, changed dataset sizes and Losses which are Cauchy and Generalized Extreme Value (Gev) losses. These newly trained models using Monoloco were observed to perform better. We evaluated these trained models on Kitti dataset and found improvised results than existing Monoloco. 

Keywords: Deep Learning, Kitti, Nuscenes, MonoLoco, Loss Function,Neural Networks.