Automatic Annotation and Segmentation of Object Instances With Deep Active Curve Network

Abstract

Object instance segmentation is a challenging task in computer vision research, and instance annotation is its key sub-task and a basic component of segmentation models. In this paper, we propose the Deep Active Curve Network (DACN) which combines powerful ResNet models with GCN-based active curves. This new method regards instance segmentation as a points/vertices, edges and object masks prediction task instead of only a pixel-labeling problem. The multi-scale encoder firstly predicts a coarse result through a combination of the edge and segmentation features in a multi-task learning framework, which is effective dealing with objects with rough boundaries. In order to generate an accurate object annotation, an iterative Graph Convolutional Network (GCN) is used to correct the encoder's feature map and move all vertices of the predicted coarse result to the edge of the corresponding ground truth. A novel weighted loss function further estimates the location of points, edges, and segmented areas, and optimizes the annotation results. Finally, we use the improved 5-interpolation Catmull-Rom spline (CRS) algorithm which exploits key points to control the active curves around objects. In the experimental analysis, we demonstrate the effectiveness of our proposed method on three datasets in automatic mode, including Cityscapes, ADE20K and Rooftop. We further show the generalization ability of our approach on two novel cross-domain datasets.

Keywords

• Object instance segmentation
• object annotation
• graph convolutional network
• multi-task learning
• active curves