Proceedings of the 6th Computer Science Research Days, JRI 2023, 18-20 December 2023, Ouagadougou, Burkina Faso

Research Article

Advancing Structured Prediction for 3D Polygonal Model Reconstruction from Single and Multiple Scene Images

Download10 downloads
  • @INPROCEEDINGS{10.4108/eai.18-12-2023.2348109,
        author={Debjyoti  Chattopadhyay and Isha  Sahni and Arpan Dutta  Chowdhury},
        title={Advancing Structured Prediction for 3D Polygonal Model Reconstruction from Single and Multiple Scene Images},
        proceedings={Proceedings of the 6th Computer Science Research Days, JRI 2023, 18-20 December 2023, Ouagadougou, Burkina Faso},
        publisher={EAI},
        proceedings_a={JRI},
        year={2024},
        month={6},
        keywords={computer vision 3d-model reconstruction multiple--view reconstruction},
        doi={10.4108/eai.18-12-2023.2348109}
    }
    
  • Debjyoti Chattopadhyay
    Isha Sahni
    Arpan Dutta Chowdhury
    Year: 2024
    Advancing Structured Prediction for 3D Polygonal Model Reconstruction from Single and Multiple Scene Images
    JRI
    EAI
    DOI: 10.4108/eai.18-12-2023.2348109
Debjyoti Chattopadhyay1,*, Isha Sahni2, Arpan Dutta Chowdhury3
  • 1: Gupdhup, Goregaon, Mumbai, Maharashtra
  • 2: IBM, Bhartiya City Bengaluru
  • 3: Tata Unistore, Mumbai, Maharashtra
*Contact email: debjyotisonuabhi@gmail.com

Abstract

We introduce a novel approach for structured prediction aimed at reconstructing 3D polygonal models from both single and multiple images capturing a scene. Building upon recent advancements in single-view reconstruction, we embrace the indoor Manhattan hypothesis category – an intricate set of potential outputs characterised by complex internal constraints – integrated into a structured prediction framework. Our methodology is adaptable for learning in both single-view and multiview scenarios. It is demonstrated that this chosen hypothesis category enables the optimization of diverse high-level loss functions, including metrics such as the relative depth error. Our achieved outcomes surpass the current state-of-the-art, showcasing an enhancement of over 50% in a specific metric.