Three dimensional shape retrieval is a fairly new concept being studied all over the world. It is notable and essential because it can be applied to multiple disciplines in society, including: computer vision, CAD models, computer graphics, molecular biology, etc. For this project, 907 2D models from the Princeton Shape Benchmark (PSB) were rendered as depth images from 20 views. The models feature 92 different classes, ranging from humans to houses. David Lowe s Scale Invariant Feature Transform (SIFT) algorithm was used in this project to normalize the images, find key points on each of the views, and create a specific feature vector to describe its respective key point. A comparison of these 907 objects was done by finding similarity between key points and their feature vectors on each of the objects corresponding views. After using the SIFT algorithm, the results were further filtered using Euclidean distance differences and spatial restrictions. By adding the spatial restrictions, it prevented the code from matching a hand to a foot; this is because the x, y, and z coordinates of the feature vectors would not be in the same general spatial area . Lastly, an equation was written to calculate the overall similarity of the objects. The different objects were then ordered based on similarity and stored in a distance matrix . The accuracy of the code and the results were evaluated by comparing the distance matrix results to the results yielded from twelve other shape descriptors (not SIFT) using the PSB base classification.
Citation: NIST Interagency/Internal Report (NISTIR) - 7625Report Number:
NIST Pub Series: NIST Interagency/Internal Report (NISTIR)
Pub Type: NIST Pubs
Three Dimensional Shape Retrieval, Scale Invariant Feature Transform, Spatial Restrictions, SIFT