Fully densely linked and strongly correlated road scene instance segmentation

Unlike conventional indoor or outdoor scenes, road images in driverless scenes usually have the characteristics of high resolution, large width (indicating a large span and containing many targets), and variable background. As a mainstream algorithm for the instance segmentation task, Polar-Mask can balance segmentation accuracy and real-time performance to some extent. However, confronted with the road scene images, its feature extraction is inadequate, and the regression branch and the classification branch in its network structure are disjoint, ignoring the potential correlation between instance contour and instance category. To overcome this shortcoming, a Polar-Mask-based fully densely linked and strongly correlated instance segmentation network (FCSIS-Polar) is proposed. Specifically, the original cascaded convolutional layers in Polar-Mask are densely connected to enhance the feature extraction of the residual network. In addition, the category features are co-encoded with the original mask prediction results as a priori information to establish a contour-category correlation. The experiment results based on the Cityscapes dataset corroborate its performance, which can achieve a segmentation accuracy of 26.4 % and a segmentation speed of 14.2 FPS even in small images.