Instance Segmentation with Separable Convolutions and Multi-level Features

Instance segmentation is a challenging task for it requires not only bounding-box of each instance but also precise segmentation mask of it. Recently proposed fully convolutional instance-aware semantic segmentation (FCIS) has done a good job in combining detection and segmentation. But FCIS cannot make use of low level features, which is proved useful in both detection and segmentation. Based on FCIS, a new model is proposed which refines the instance masks with features of all levels. In the proposed method, large kernel separable convolutions are employed in the detection branch to get more accurate bounding-boxes. Simultaneously, a segmentation module containing boundary refinement operation is designed to get more precise masks. Moreover, the low level, medium level, and high level features in Resnet-101 are combined into new features of four different levels, each of which is employed to generate a mask of an instance. These masks are added and refined to produce the final most accurate one. With the three improvements, the proposed approach significantly outperforms baseline FCIS as it provides 4.9% increase in mAPr@0.5 and 5.8% increase in mAPr @0.7 on PASCAL VOC. © Copyright 2019, Institute of Software, the Chinese Academy of Sciences. All rights reserved.