Hypergraph Induced Graph Convolutional Network for Multi-Label Image Recognition

Abstract

It is well known that exploiting label connections is important for multi-label image recognition. However, many of the existing methods utilize the connections between label pairs, ignoring high-order connections, thus may result in the degradation of recognition performance. A few methods can capture high-order label connections but have high complexity and low scalability. Inspired by the nature of hypergraph coding high-order connections, we propose a Hypergraph induced Graph Convolutional Network (HI-GCN), which can capture high-order label connections with high adaptivity and scalability. Specifically, we first build an adaptive hypergraph on labels in a data-driven manner, which allows the high-order label connections to be exploited adaptively and scalability. Then, by updating label embeddings via using Hypergraph induced Graph Convolutional network (HI-GCN), the label embeddings are mapped to label classifiers with high-order connections. Experimental results on MS-COCO, SUN and ESP datasets validate the effectiveness of our approach.