MDNet: A Multi-image Input Real-Time Semantic Segmentation Model with Decoupled Supervision

Current state-of-the-art Real-Time Semantic Segmentation Model is still not fast enough. They spend too much time on processing images in a deep CNN to grab the spatial and context information. Somehow, this information may not be so deterministic. In this work, we come up with a multi-image input real-time semantic segmentation model with decoupled label supervision. It can decrease the computational time and keep a relatively high precision of semantic segmentation meanwhile. The novelty of our model lies is picking up the decoupled label supervision to be our loss function and combining it with a multi-branch image processing framework. The edge detection module can not only improve the recognition of the differences between object body and edge but also guarantee the processing procedure of our network to be faster enough. Apart from this, the multi-branch image processing framework is not a burden of running time. Our network is trained on difficult datasets like CamVid and has favourable quality in real-time testing. The mean class IoU of our network is 66.6. It is the highest one among all of the other comparisons.