Building Better Models: A Benchmark on Feature Extractors and Matchers for Structure from Motion in Construction Sites

Abstract

The increased popularity of Structure from Motion (SfM) techniques has revolutionized 3D reconstruction in various fields, including construction site mapping. SfM enables the generation of detailed and accurate 3D models from real-world scenes captured in 2D images, facilitating better project monitoring, analysis, and decision-making. Key to reconstruction using SfM, is the feature extraction and matching process, which identifies and matches corresponding points in different images to reconstruct the scene accurately. Benchmarks have been conducted to evaluate the performance of traditional and learning-based feature extraction and matching. However, these studies have not focused on construction site mapping and predominantly evaluate single components of the SfM pipeline. This thesis aims to provide a comprehensive evaluation of traditional and learning-based feature extraction and matching methods within the SfM pipeline, focusing on the reconstruction quality and their effectiveness in construction site mapping. Traditional feature extraction methods like SIFT, AKAZE, and ORB, are compared against advanced learning-based methods like SuperPoint, D2-Net, DISK, SOS-Net, and R2D2. Similarly, matching techniques like SuperGlue and LightGlue are compared against traditional ones such as brute-force and nearest neighbor. Results indicate that learning-based methods generally outperform traditional approaches in terms of robustness and accuracy, particularly in scenarios with complex lighting and limited visual overlap. Deep learning methods also demonstrated superior feature extraction and matching capabilities, producing more detailed and accurate reconstructions with fewer missing areas, as well as, faster processing times thanks to their GPU-accelerated implementations. The findings underscore the importance of selecting appropriate techniques based on specific scene characteristics and desired outcomes when performing construction site mapping using SfM, and the potential of learning-based methods to enhance the quality and efficiency of 3D reconstruction in this domain. Likewise, the results highlight the need for further research on refining these methods to handle more complex real-world scenarios effectively, improving their robustness and computational efficiency for broader practical adoption.