Spatial Front-Back Relationship Recognition Based on Partition Sorting Network

The current development in computer vision highlights the significance of comprehending the semantic features of images. However, the spatial front-back relationships between objects, which constitute a fundamental semantic feature, have received limited attention. To address this gap, we propose a novel neural network termed Spatial Front-Back Relationship Partition Sorting Network (SFBR-PSortNet), specifically designed for recognizing spatial front-back relationships among objects within images. SFBR-PSortNet is an end-to-end deep convolutional neural network that systematically generates a set of triples representing the spatial front-back relationships between every pair of objects in an input image. The key technical innovations of SFBR-PSortNet include the utilization of bottom keypoints of objects, which serve a dual purpose of enabling object category recognition and providing implicit depth information to enhance spatial front-back relationship reasoning, and the introduction of a Partition Sorting mechanism to construct a comprehensive spatial front-back relationship graph among all objects. Extensive experiments conducted on data derived from the KITTI dataset demonstrate the effectiveness of our network for spatial front-back relationship recognition, achieving a Precision of 0.876 and a Recall of 0.856, respectively. The results validate the practical applicability and robustness of our network in real-world road scenarios, underscoring its potential to enhance the accuracy of computer vision systems in complex environments.