Two-dimensional spatial orientation relation recognition between image objects

Abstract

Recent advances in computer vision have concentrated on comprehension of the semantic features of images, particularly the spatial relations between objects—a fundamental semantic feature of visual scene understanding. This study systematically addresses the recognition problem of two-dimensional spatial orientation relations and develops the Target Spatial Orientation Vector Field (TSOVF) algorithm, a novel end-to-end framework to explicitly model spatial orientation dependencies. TSOVF algorithm introduces the learnable spatial orientation vector field to effectively encode the spatial orientation relation into a deep convolutional neural network model. The proposed architecture features a dual-branch design: the T-branch identifies object central points and classifies categories via keypoint estimation, while the S-branch constructs a pixel-level spatial orientation vector field. Each vector in this field quantifies the angular orientation between object pairs, with aggregated vector data determining the final spatial relation category. A dedicated fusion module synthesizes features from both branches, generating a structured triple list that documents detected objects, their inter-object spatial orientations, and associated confidence scores. Evaluated on a PASCAL VOC2012-derived dataset, TSOVF algorithm achieves 94.8 % global accuracy and a class-balanced geometric mean (G-mean) of 0.798, demonstrating robust performance across various spatial configurations. For dominant orientation categories, the algorithm attains up to 95.9 % precision and 94.7 % F1-score, establishing it as a foundational benchmark for spatial relation recognition. These results validate TSOVF’s capacity to advance fine-grained visual relationship detection while providing a reproducible framework for future research in spatial-semantic analysis.