A Comprehensive Survey on 3D Single-View Object Reconstruction.

Area of Science:

• Computer Vision
• 3D Graphics
• Machine Learning

Background:

• Single-view 3D object reconstruction (SVOR) faces challenges due to incomplete image data, limited 3D annotations, and object shape variability.
• Recent progress in novel view synthesis (NVS) has significantly advanced SVOR, yet comprehensive reviews are lacking.

Purpose of the Study:

• To review the latest advancements in SVOR, with a specific focus on NVS-based methods.
• To address the limitations of current SVOR evaluation metrics that do not align with human visual perception.
• To propose a more accurate and perceptually relevant evaluation methodology for SVOR.

Main Methods:

• A comprehensive review of recent NVS-based approaches in single-view 3D object reconstruction.
• Development of a weighted evaluation method that incorporates part saliency to better reflect human visual perception.
• Introduction of a novel technique for automatic perception of reconstruction discrepancies.

Main Results:

• The study highlights significant progress in SVOR driven by NVS techniques.
• The proposed weighted evaluation method demonstrates improved accuracy and consistency compared to existing metrics.
• The novel discrepancy perception technique aids in identifying critical reconstruction failures.

Conclusions:

• Existing SVOR reviews need updating to include NVS-based advancements.
• The proposed evaluation method offers a more human-aligned assessment of 3D reconstruction quality.
• This work provides valuable insights and methodologies for both research and practical applications in SVOR.