3D-to-2D embedding-driven object classification with multiview robustness
Accurate recognition of objects from 2D images is essential in applications such as robotics, augmented reality (AR), and digital twin systems, where visual understanding directly influences spatial interaction and decision-making. This work presents a 3D-to-2D embedding-based learning framework designed to bridge the gap between synthetic and realworld data for reliable object classification and pose estimation. In this framework, a large collection of 3D object models is rendered from multiple viewpoints to capture the spatial variability encountered in real scenarios. Each rendered image is passed through a pretrained ResNet50 network to generate high-dimensional embeddings that encode shape and appearance features invariant to lighting and orientation. These embeddings are then applied to real-world 2D datasets, allowing the model to classify real objects with minimal real-world annotations. Experimental results show that the proposed method generalizes well and achieves high classification accuracy across various real-image benchmarks, demonstrating the benefit of learning from synthetic 3D data. Additionally, orientation regression is incorporated to predict azimuth, elevation, and roll angles, providing a foundation for pose-aware recognition critical for AR visualization and robotic manipulation. While current pose estimations have limited precision due to domain differences between synthetic and real data, the findings suggest promising avenues for future improvements, including domain adaptation, contrastive embedding alignment, and multi-view fusion strategies to enhance spatial consistency and facilitate real-time deployment in AR and digital twin applications.