Efficient Instance Segmentation Network

Abstract

We present an efficient, flexible, fast and accurate framework for real-time instance segmentation. We call it Efficient Instance Segmentation Network, denoted as EISNET. Our method is motivated by the Mask R-CNN and YOLACT. Mask R-CNN enables instance segmentation by adding an extra branch at the Faster R-CNN framework to produce mask for each object. Due to limitation of the inefficiency of two stage detector, Mask R-CNN is not suitable for real-time scene. We therefore propose EISNET which enables instance segmentation by adding two branches to the one-stage detector-RetinaNet. We call it Efficient since we use modified EfficientNet as the backbone of our framework, which results in high accuracy with few parameters and FLOPS. In addition, we provide a modified bi-directional FPN (Feature Pyramid Network) module, which thus allows efficient multi-scale feature fusion. Given the credit to these design techniques, our EISNET achieves 31.2 mAP with only 17.2M parameters and 3.5B FLOPS on the COCO dataset. More significantly, our model can achieve more than 35 FPS on single 1080Ti GPU, which fits the most real-time requirements. With a better GPU, we could even achieve higher mAP while keeping the real-time property of more than 30 FPS.