Wyniki wyszukiwania - BazTech

Ograniczanie wyników

Znaleziono wyników: 2

Liczba wyników na stronie

Wyniki wyszukiwania

Wyszukiwano:
w słowach kluczowych: distance metrics

Sortuj według:

Ogranicz wyniki do:

The comparison of distances metrics in descriptor matching methods utilised in TLS-SfM point cloud registration

Markiewicz Jakub

Reports on Geodesy and Geoinformatics

2025

Vol. 119

39--61

Advanced measurement techniques, such as Terrestrial Laser Scanning (TLS), play a vital role in documenting cultural heritage and civil engineering structures. A key aspect of these applications is the accurate registration of point clouds. Conventional TLS methods often rely on manual or semi-automated correspondence detection, which can be inefficient for large or complex objects. Structure-from-Motion Terrestrial Laser Scanning (SfM-TLS) offers an alternative methodology, comprising two primary phases: correspondence search and incremental reconstruction. Descriptor matching in SfM-TLS typically employs the L2 norm to measure Euclidean distances between features, valued for its simplicity and compatibility with algorithms like SIFT. This study investigates the influence of various distance metrics on descriptor matching during the correspondence search stage of SfM-TLS. Eight metrics were analysed: Bray-Curtis, Canberra, Correlation, Cosine, L1, L2, Squared Euclidean, and Standardised Euclidean. Synthetic data experiments highlighted challenges in keypoint detection and matching due to measurement angles, material characteristics, and 3D-to-2D transformations. Simulations incorporating Gaussian noise demonstrated that image rotation and skew significantly affected tie-point accuracy, more so than variations in intensity. In field applications involving cultural heritage sites and building interiors, the L1 and Squared Euclidean metrics yielded higher accuracy, while the Canberra metric underperformed. Metric selection was found to have a greater impact on complex geometries, such as historical structures, compared to simpler forms. Consequently, this study recommends the L1 and Squared Euclidean metrics for pairwise SfM-TLS registration, as they exhibit robustness in maintaining high accuracy and completeness across a variety of architectural scenarios.

Human Iris Classification through Histogram of Oriented Gradient Features with Various Distance Metrics

Mukherjee Arnab, Islam Zahidul, Ali Lasker Ershad

Machine Graphics and Vision

2024

Vol. 33, No. 3/4

97--124

Human iris classification remains an active research area in the fields of biometrics as well as computer vision. In iris biometrics, most of the visible or near-infrared (NIR) eye images suffer from multiple noise sources, and the dispersive spectrum changes hugely. These changes occur due to spattering, albedo, and spectrum absorbance selectively. However, accurate iris classification for distance images is still a challenging task. To solve it effectively, we propose a machine learning (ML)-based iris classification employing a dense feature extraction method with various distance metrics. More specifically, this learning model focuses on the Histogram of Oriented Gradients (HOG) descriptor and K-Nearest Neighbour (K-NN) classifier with various distance metrics. The HOG descriptor has some advantages for this proposed distant-based iris classification, for example, insensitive to multiple lighting and noises, shift invariance, capacity to tolerate iris variations within the classes, etc. Additionally, this study investigates the most reliable distance metric that is less affected by different levels of noise. A publicly accessible CASIA-V4 distance image database is conducted for the experimental evaluation. To evaluate the performance of the classification models, we consider different measures such as recall, precision, F₁-score, and accuracy. The reported results are tabulated as well as optimized through Receiver Operating Characteristic (ROC) curves. The experimental results demonstrate that the Canberra distance metric with low dimensional HOG features provides better recognition accuracy (90.55%) compared to other distance metrics.