Efficient Semantic Segmentation through Dense Upscaling Convolutions

Semantic segmentation is the classification of each pixel in an image to an object, the resultant pixel map has significant usage in many fields. Some fields where this technology is being actively researched is in medicine, agriculture and robotics. For uses where the resources or power requirements are restricted such as robotics or where large amounts of images are required to process, efficiency can be key to the feasibility of a technique. Other applications that require real-time processing have a need for fast and efficient methods, especially where collision avoidance or safety may be involved. We take a combination of existing semantic segmentation methods and improve upon the efficiency by the replacement of the decoder network in ERFNet with a method based upon Dense Upscaling Convolutions, we then add a novel layer that allows the fine tuning of the decoder channel depth and therefore the efficiency of the network. Our proposed modification achieves 20-30% improvement in efficiency on moderate hardware (Nvidia GTX 960) over the original ERFNET and an additional 10% efficiency over the original Dense Upscaling Convolution. We perform a series of experiments to determine viable hyperparameters for the modification and measure the efficiency and accuracy over a range of image sizes, proving the viability of our approach.