More Competitive Feature Extraction Network for Instance Segmentation

Instance segmentation has a wide range of applications in autonomous driving, video security and so on. Mask R-CNN, which introduces Feature Pyramid Networks (FPN), is a simple and effective instance segmentation framework. However, it still has some problems such as false detection, missed detection, and low instance segmentation accuracy. To address such problems, we make corresponding improvements on Mask R-CNN. A more competitive feature extraction module, which we call Squeeze-and-Excitation feature model, is thereupon proposed. Squeeze-and-Excitation feature model performs a bottom-up feature fusion on the FPN output layer, making the underlying features easier to propagate. In particular, a concurrent spatial and channel Squeeze-and-Excitation module (scSE) is employed in it. The application of scSE can improve the adaptability of the feature channel and aggregate relevant spatial information, which greatly reduce false detection and missed detection. In addition, during the final segmentation phase, a spatial Squeeze-and-Excitation module (sSE) is used to refine the segmentation. The experimental result reveals that, both detection and instance segmentation have achieved significant improvements in the MS COCO dataset.

I. Introduction

In recent years, deep neural networks has brought rapid development in computer vision field. The prediction ability of deep networks is gradually developing from image level prediction to regional prediction, pixel prediction and instance prediction. Instance segmentation has both practicality and challenge, and has become a research hotspot in computer vision. Instance segmentation not only distinguishes between object categories in the foreground, but also different instances of the same category.