DEL: Deep Embedding Learning for Efficient Image Segmentation

Yun Liu¹, Peng-Tao Jiang¹, Vahan Petrosyan², Shi-Jie Li¹, Jiawang Bian³, Le Zhang⁴, Ming-Ming Cheng¹

¹Nankai University        ²KTH Royal Institute of Technology        ³University of Adelaide        ⁴Advanced Digital Sciences Center

Abstract

Image segmentation has been explored for many years and still remains a crucial vision problem. Some efficient or accurate segmentation algorithms have been widely used in many vision applications. However, it is difficult to design a both efficient and accurate image segmenter. In this paper, we propose a novel method called DEL (deep embedding learning) which can efficiently transform superpixels into image segmentation. Starting with the SLIC superpixels, we train a fully convolutional network to learn the feature embedding space for each superpixel. The learned feature embedding corresponds to a similarity measure that measures the similarity between two adjacent superpixels. With the deep similarities, we can directly merge the superpixels into large segments. The evaluation results on BSDS500 and PASCAL Context demonstrate that our approach achieves a good trade-off between efficiency and effectiveness. Specifically, our DEL algorithm can achieve comparable segments when compared with MCG but is much faster than it, i.e. 11.4fps vs. 0.07fps.

Papers

• DEL: Deep Embedding Learning for Efficient Image Segmentation, Yun Liu, Peng-Tao Jiang, Vahan Petrosyan, Shi-Jie Li, Jiawang Bian, Le Zhang, and Ming-Ming Cheng, IJCAI-ECAI, 2018. [Project page] [pdf] [bib] [Source code]

We have released the code and data for plotting the edge PR curves of many existing image segmentation and edge detection methods here.

Algorithm Pipeline

Network Architecture

Superpixel Merging

Evaluation on BSDS500 Dataset

Evaluation on PASCAL Context Dataset