Research on Scene Parsing Algorithm Cascading Object Detection Network

摘要

Using deep learning for visual scene parsing will satisfy the demand of the next generation of automatic driving technology. However, current parsing algorithms are not mature enough for practical applications unless higher accuracy and efficiency are obtained. We propose a novel scene parsing algorithm framework which integrates the object detection technologies into convolution neural network to improve the overall effectiveness. The framework consists of three components: i) a scene parsing network presenting primary semantic segmentation result. ii)an object detection network calculating the location and confidence of the targets in images. iii) an integration and filter module that cascades previous two results. Extensive experiments suggest that our model is capable of practical use and achieving more favorable scene parsing performance of mIoU score as 69.4% on CamVid dataset.