Content-Based Retrieval of 3D Head Models Using a Single Face View Query

INVESTIGATORS

Yeung Pui Fong, Dr. Hau San Wong and Prof. Horace H S Ip

BRIEF DESCRIPTION

This project proposes a new approach for 3D human head model retrieval using a single 2D face view only. In this way, the query can be conveniently specified in the form of a single portrait which is in most cases readily available, instead of a 3D model query which is difficult to construct, or text-based query which in general cannot describe the models adequately. To realize this approach, a mapping between the 2D face views and the 3D models needs to be established. In our case, 3D models are represented with a set of adaptive basis functions, while their corresponding 2D face views are characterized with a set of eigenface basis functions. In this way, a particular model and its associated face view can be identified by two separate set of expansion coefficients. To associate the two, we propose to exploit neural network techniques to identify a mapping. With this 2D-3D mapping, we can thus estimate a set of associated 3D expansion coefficients for the input query to retrieve the relevant models in the database.

REPRESENTATIVE MODELS FROM DATABASE

Class Representative Models(Female) 1 2 3 4 5

Class Representative Models (Male) 1 2 3 4 5

Representatives models from our 3D head model database

SAMPLE RESULTS

First ten retrieved models of two input queries

Precision recall characterization of retrieval performance

PUBLICATION

• Yeung Pui Fong, Dr. Hau San Wong and Prof. Horace H S Ip, "Content-Based Retrieval of 3D Head Models Using a Single Face View Query", Pacific-Rim Conference on Multimedia 2004, Tokyo, Japan, Nov 30 - Dec 3 2004