No boundary left behind in semantic segmentation

• Jefferson Fontinele Da Silva UFMA
• Bernardo Silva UFBA
• Luciano Oliveira UFBA

Resumo

This paper proposes a novel network architecture for image semantic segmentation based on attention mechanisms placed on specific points inside a convolutional neural network. Attention is explored across our network to integrate information from object boundary and a baseline semantic segmenter (inner segmentation). We call our novel network Attention-fitted Fusion of boundary and Inner Segmentation (AFIS), which combines the two streams through a set of attention gates, forming an end-to-end network. We performed an extensive evaluation of our method over four public challenging data sets (Cityscapes, CamVid, Pascal Context, and Mapillary Vistas), finding superior results when compared with other twelve state-of-the-art segmenters, considering the same training conditions.

Palavras-chave: Training, Graphics, Semantic segmentation, Semantics, Network architecture, Logic gates, Convolutional neural networks