Image-Based 3D Object Reconstruction: State-of-the-Art and Trends in the Deep Learning Era.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning
• Deep Learning

Background:

• 3D reconstruction is a complex, ill-posed problem studied for decades.
• Image-based 3D reconstruction using convolutional neural networks (CNNs) has shown significant progress since 2015.
• Deep learning techniques are revolutionizing the estimation of 3D shapes from images.

Purpose of the Study:

• To provide a comprehensive survey of recent developments in deep learning-based 3D reconstruction.
• To focus on methods using CNNs for estimating 3D shapes of generic objects from RGB images.
• To review techniques for both single and multiple image inputs.

Main Methods:

• Literature organized by shape representations, network architectures, and training mechanisms.
• Analysis of works employing deep learning for 3D shape estimation.
• Inclusion of methods reconstructing generic objects, human bodies, and faces.

Main Results:

• Demonstration of impressive performance by CNN-based 3D reconstruction methods.
• Categorization of recent research based on key technical aspects.
• Identification and comparison of performance across key studies.

Conclusions:

• Deep learning, particularly CNNs, has significantly advanced image-based 3D reconstruction.
• The field is rapidly evolving with diverse approaches to shape representation and network design.
• Open problems and future research directions in 3D reconstruction are highlighted.