Pinhole Tabanlı Sentetik Veriyle Eğitilmiş MLP ile Özel Tasarım Etiket Tabanlı Poz ve Mesafe Tahmini
Year: 2026
Type: conference-paper
Abstract: This study focuses on estimating the pose and distance information of objects using a custom-designed tag. In this context, this study presents a multi-output Multilayer Perceptron (MLP) model trained using pinhole-based synthetic measurement data. In the proposed method, synthetic measurement data covering different distance, image-center, and orientation conditions were generated using effective pinhole parameters consistent with the real camera geometry. A 19-dimensional geometric feature set was extracted from the center positions of the tag sub-components and used as the input data of the MLP model. The model architecture and training hyperparameters were systematically determined through an Optuna-based search process. On the independent synthetic test set, an overall R2 value of 0.9873 was obtained, suggesting that the model exhibited consistent performance in the multi-output estimation task. Real-environment tests showed that ROI determination, subcomponent detection, geometric feature extraction, MLP-based estimation, and pinhole-based reprojection could be integrated and executed on video frames. The obtained results show that the method provides an applicable framework for visual localization tasks in areas such as robotic manipulation, augmented reality, and human-machine interaction. As future work, alternative regression strategies are planned to be investigated with the aim of reducing deviations, particularly in yaw and pitch estimation.
Author: Emre Erkan
Institution: Batman University
Language: Turkish
Topic: Advanced Vision and Imaging
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7223658898