Mobile-Based 3D Modeling: An In-Depth Evaluation for the Application in Indoor Scenarios

Area of Science:

• Computer Vision
• 3D Reconstruction
• Machine Learning

Background:

• Indoor environment modeling is crucial for augmented, virtual, and extended reality applications.
• Digital transformation drives the need for detailed 3D environment models and surface mapping for virtual element integration.
• Existing methods focus on generating navigable 3D models and inserting virtual objects into real scenes.

Purpose of the Study:

• To review state-of-the-art (SoA) learning-based methods for 3D scene reconstruction using structure from motion (SFM).
• To evaluate the generalization capability of a recent SoA network on unseen indoor environment data.
• To assess depth map prediction accuracy using the absolute relative (AbsRel) metric.

Main Methods:

• Review of learning-based 3D scene reconstruction techniques.
• Implementation and evaluation of a recent SoA network for depth map and camera pose prediction from video streams.
• Quantitative analysis using the absolute relative (AbsRel) error metric.

Main Results:

• The paper provides a comprehensive overview of current learning-based SFM approaches for indoor environments.
• The evaluation highlights the performance and generalization abilities of a selected SoA network.
• Depth map prediction accuracy is quantified using the AbsRel metric, serving as a baseline for comparison.

Conclusions:

• Learning-based methods show promise for accurate 3D indoor environment reconstruction.
• Generalization to unseen data is a critical factor in evaluating the robustness of these methods.
• The study establishes a benchmark for future research in indoor scene understanding and modeling.