Failure Detection in Pedestrian Detection : A Fully Convolutional Neural Network approach
Author
Apostolopoulos, Christos
Term
4. term
Publication year
2016
Submitted on
2016-06-02
Pages
85
Abstract
Pedestrian detection forms a key problem in computer vision, with applications that can greatly affect every day life. In this project, failures in pedestrian detectors are attempted to be refined by re-evaluating their results via a fully convolutional neural network. The network is trained on a number of datasets which include a custom partial occluded pedestrian dataset. The networks efficiency was evaluated by examining the loss through training (iterations) and by measuring the mean intersection union across classes. As a detector, accuracy was measured by comparing the network given a detectors result against state-of-the-art but also by measuring the networks refined result against the detector without the network. It was found that although results vary, the proposed network shows promising results.
Documents