Deep learning-based car detection

Abstract

Many of the recent state-of-the-art object detection performances in computer vision evolved around deep learning. Faster R-CNN being one of the most recent breakthroughs that harnesses the power of convolutional neural networks (CNNs) to extract features was able to achieve ground-breaking detection results on general objects. This work applied Faster R-CNN in one specific context that is vehicle detection, and explored several methods to adapt and optimize the performance of Faster R-CNN on bus and car object classes. The methods that were experimented with include modifying anchor ratios at the region proposal stage and augmenting training data with other supervised datasets as well as bootstrapped web images. The resulting models were able to outperform baselines by up to 5-10% on the two vehicle classes. In addition, an interactive web demo was created to deploy the resulting model and accept images for detection.