This repository contains the code used during the master thesis project "Vehicle detection and Road Scene Segmentation using Deep Learning" (Link to report). All approaches are based on Convolutional Neural Networks (CNNs) and are implemented in LuaJIT using Torch7.
All implementations uses cutorch -- a CUDA backend for Torch7, and cudnn -- a GPU-accelerated library for deep neural networks.
More specifically, the repository includes;
-
An MNIST-based detection implementation. A CNN is trained to classify regions of an image with spread out MNIST digits. Bounding boxes for digits are achieved through regression. The task was implemented as a proof if concept for the "detection through classification" approach used for vehicle detection.
-
A vehicle detection implementation trained on the KITTI object detection data set. This was the main focus for the thesis.
-
An implementation of a network for semantic segmentation based on a deconvolutional approach by Noh et al. The network is trained on the Cityscapes data set.
The documentation of these implementations is very breif and meant to give a quick overview of the structure to be used for similar projects rather than a useful product in it self.