Visual Relationship Detection Based on Local Feature and Context Feature

Visual Relationship Detection(VRD) is one of the key tasks in computer vision and plays an important role in image understanding. The aim of VRD is to detect the relationship between objects in the input image and then to generate a number of relationship triples, namely <;subject, predicate, object>. Previous works on this task usually used the union box of the subject and object to predict the predicate by Convolution Neural Network(CNN). However only the local information in the image is not enough to predict effectively. So we combined the local information, the context information and statistical dependency together into an end-to-end model to do VRD task. Experiments shows that our methods can improve the precision of prediction.