GERRS: Removing Ghost Effects from Real-World Scenarios in 3D Pose Estimation via Zero-shot Inference Approach

In the realm of computer vision, 3D pose estimation algorithms have achieved remarkable proficiency in controlled benchmark datasets. However, their performance often falters when confronted with the complexities of real-world monocular video scenarios. One of the most formidable challenges in this context is the presence of intricate and unpredictable backgrounds. These backgrounds, due to their uncertain textures, are frequently misinterpreted as human subjects, giving rise to the persistent and vexing 'ghost effect.' This phenomenon not only distorts the accuracy of pose estimation but also hinders its applicability in various practical settings. In response to this conundrum, we present GERRS, a plug-in-play and innovative zero-shot technique designed to address the 'ghost effect' problem by effectively managing uncertain textures. GERRS integrates four core components: moving human detection, motion detection, background texture pre-processing, and 3D pose estimation. This comprehensive approach has been subjected to rigorous validation, which has conclusively demonstrated its unparalleled efficacy in mitigating the ghost effect. Our method brings a remarkable enhancement to the accuracy of 3D pose estimation in real-world monocular RGB videos, marking a significant stride towards the realization of robust and reliable pose estimation in dynamic and unpredictable environments. The proposed approach can be incorporated with any existing 3D pose estimation approaches to enhance the performance of the pretrained approach. By tackling the 'ghost effect' head-on, GERRS holds the promise of revolutionizing the field of computer vision, opening doors to more precise and dependable applications in a wide range of domains, from augmented reality to robotics.