Abstract
Semantic image segmentation aims to assign a semantic label to each pixel in an image. Recent state-of-the-art approaches are mainly based on Convolutional Neural Networks. Although these approaches achieve outstanding performance, they adopt very complex CNN models. As the result, they usually require larger training dataset and spend more time on both training and inference stages. In contrast to recent complex CNN-based approaches, we propose to simplify an existing CNN architecture, VGG-16, but do not compromise the segmentation performance. Firstly, we propose a basic model by replacing the original fully-connected layers with several convolutional and pooling layers for extracting hierarchical features. We then use the extracted hierarchical features to generate multi-scale predictions, and aggregate all predictions to derive one dense prediction result. Furthermore, we extend the basic model with cross-layer feature concatenation to jointly exploit the information from lower- and higher-level layers. Experimental results show that with only one-fourth the parameters of the original VGG and no post-processing or Conditional Random Field refinement, the proposed model achieves comparable results on three popular datasets: SIFT Flow, Pascal VOC 2012, and Pascal Context.