Human–Object Interaction Detection with Ratio-Transformer

Abstract

Human–object interaction (HOI) is a human-centered object detection task that aims to identify the interactions between persons and objects in an image. Previous end-to-end methods have used the attention mechanism of a transformer to spontaneously identify the associations between persons and objects in an image, which effectively improved detection accuracy; however, a transformer can increase computational demands and slow down detection processes. In addition, the end-to-end method can result in asymmetry between foreground and background information. The foreground data may be significantly less than the background data, while the latter consumes more computational resources without significantly improving detection accuracy. Therefore, we proposed an input-controlled transformer, “ratio-transformer” to solve an HOI task, which could not only limit the amount of information in the input transformer by setting a sampling ratio, but also significantly reduced the computational demands while ensuring detection accuracy. The ratio-transformer consisted of a sampling module and a transformer network. The sampling module divided the input feature map into foreground versus background features. The irrelevant background features were a pooling sampler, which were then fused with the foreground features as input data for the transformer. As a result, the valid data input into the Transformer network remained constant, while irrelevant information was significantly reduced, which maintained the foreground and background information symmetry. The proposed network was able to learn the feature information of the target itself and the association features between persons and objects so it could query to obtain the complete HOI interaction triplet. The experiments on the VCOCO dataset showed that the proposed method reduced the computational demand of the transformer by 57% without any loss of accuracy, as compared to other current HOI methods.