• Vinay Magar
• Vaibhav Wasamkar
• Ashish Ranjan Magar
• Ravi Khatri

3D reconstruction plays a crucial role in computer vision, yet traditional Multi-View Stereo (MVS) methods struggle with unseen objects due to their dependence on large labeled datasets and dense multi-view coverage. This paper proposes a Hybrid Self-Learning 3D Reconstruction framework, integrating Few-Shot Learning (FSL) and Self-Supervised Learning (SSL) to refine depth estimation and improve generalization. By retraining Intel DPT-Large and leveraging open-source datasets, our model adapts to novel objects with minimal supervision. Experimental results show improved depth accuracy, reduced computational costs, and enhanced object reconstruction, making it a scalable solution for robotics, AR/VR, and autonomous systems. A cross-dataset testing protocol has been added to enhance generalization. We formalize FSL adaptation bounds using the PAC-learning framework to provide a rigorous theoretical foundation.

3D Model Reconstruction, GPU Acceleration, Deep Learning, Computer Vision, Parallel Computing.

"Self-Learning 3D Reconstruction Using MVS and Few-Shot Learning for Unseen Objects", International Journal of Emerging Technologies and Innovative Research (www.jetir.org), ISSN:2349-5162, Vol.12, Issue 4, page no.l83-l90, April-2025, Available :http://www.jetir.org/papers/JETIR2504B12.pdf

"Self-Learning 3D Reconstruction Using MVS and Few-Shot Learning for Unseen Objects", International Journal of Emerging Technologies and Innovative Research (www.jetir.org | UGC and issn Approved), ISSN:2349-5162, Vol.12, Issue 4, page no. ppl83-l90, April-2025, Available at : http://www.jetir.org/papers/JETIR2504B12.pdf