Ensemble Deep Learning for Human-Object Interaction Detection
Ahmed E. Mansour, Ammar Mohammed, Hussein A. Elsayed, Salwa H. El Ramly
Abstract
Human-object interaction (HOI) detection is the task of predicting the visual relationships between humans and their surroundings in images and videos by locating and inferring the interactions between human-object pairs. The majority of existing models have approached this task by detecting human and object instances and predicting interactions between them using a single model and relying on visual features to differentiate between different actions. In this article, we propose a novel method for solving the HOI detection problem by employing ensemble voting on simple models that use only spatial features to predict human-object pair actions. Additionally, we solve the false-positive pairs generated by mis-grouping and non-interaction objects in the image using a bipartite matching. Our proposed approach outperforms many state-of-the-art models on the V-COCO dataset while requiring less inference time.