MetricMask: Single category instance segmentation by metric learning

In this paper, we introduce a novel single category instance segmentation method, termed MetricMask, which can be used by easily embedding it into most off-the-shelf detection and segmentation methods. Currently, instance segmentation methods usually segment each instance after completing the object detection. Our method can implement segmentation for all instances at once by fusing object detection, semantic segmentation, and metric learning. The contributions are threefold, 1. The instance segmentation is converted to three parallel tasks, including bounding box regression, mask regression, and embedding vector of pixel regression. 2. Random sampling metric loss is proposed to optimize embedding vectors, saving GPU memory without losing instance segmentation accuracy. 3. Based on the model output(bounding box, embedding vector of pixel, and the category-level mask), we propose a Metric Operation to segment each instance. Finally, we conduct experiments on two standard datasets, including the COCO person dataset and the ISOD dataset. The experiment results show that our method has excellent competitiveness with other methods. In particular, our method has an excellent performance on the large object and outperforms the existing state-of-the-art competitors on the ISOD dataset. We hope the MetricMask of our proposed will provide a new method for instance segmentation. (c) 2022 Elsevier B.V. All rights reserved.

Automated summary

In this paper, a novel single category instance segmentation method called MetricMask is introduced, which can be easily embedded into most off-the-shelf detection and segmentation methods. By fusing object detection, semantic segmentation, and metric learning, our method can perform segmentation for all instances at once. Experimental results demonstrate that our method has excellent competitiveness on two standard datasets.