A Performance Evaluation of Local Features for Image-Based 3D Reconstruction.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Machine Learning

Background:

• Local features are crucial for image-based 3D reconstruction.
• Evaluating state-of-the-art local features is essential for advancing the field.
• Both handcrafted and machine learning-based features require thorough comparison.

Purpose of the Study:

• To conduct a comprehensive comparative evaluation of local features for 3D reconstruction.
• To assess both float and binary local features.
• To analyze performance across controlled and large-scale, unconstrained datasets.

Main Methods:

• Evaluated recently developed machine learning features and handcrafted features.
• Included both float and binary feature types.
• Utilized two datasets: one with groundtruth for quantitative analysis in controlled scenes, and internet-scale landmark datasets for qualitative analysis in unconstrained environments.

Main Results:

• Binary features provide efficient 3D reconstruction for controlled scenes, significantly reducing processing time.
• Float features demonstrate a clear advantage in large-scale datasets with numerous distracting images.
• Scale-Invariant Feature Transform (SIFT) exhibits high stability across scene types and competitive reconstruction results.

Conclusions:

• Binary features are suitable for controlled 3D reconstruction scenarios due to their speed.
• Float features are more robust for complex, large-scale datasets.
• While learned binary features lag behind handcrafted ones, CNN-based float feature learning shows future potential for 3D reconstruction.