SwinRelTR: an efficient single-stage scene graph generation model for low-resolution images

Abstract

Targeting low-resolution imagery is crucial in democratising computer vision technologies, facilitating applicability in resource-limited environments where high-resolution data is often unreachable. Scene graphs have proven to be a powerful representation for capturing the hierarchical relationships between objects in an input image, providing a structured visual scene understanding. Nevertheless, all scene graph generation models focus on high-resolution images, neglecting the challenges posed by low-resolution images. This paper presents a novel approach called SwinRelTR for generating scene graphs designed specifically for low-resolution images. The proposed model addresses the limitations associated with low-resolution images by utilising the Swin transformer as a backbone instead of the convolution neural network in the original RelTR model. The Visual Genome dataset is utilised to compare the SwinRelTR results with the state-of-the-art approaches. It has been proven that this approach outperforms several state-of-the-art approaches as well as the original RelTR model on low-resolution images.