3D Face Recognition on Depth Video
(Project done as part of summer Internship at University of Southern California (USC) under the supervision of Prof. Gerard Medioni)
The goal of this project was to develop an online system for 3D face recognition using depth videos obtained from RGB-D sensors. The work consists of two parts: - Enrollment and Recognition. Given a face detection module, in the enrollment part subject’s 3D face data was to be saved in .txt file format in the database (to be used later during recognition). Recognition module has two parts: -
- Registration of online captured data (called model) with the data stored in database. For this we used GPU based implementation of the ICP (iterative closest point) algorithm.
- Computation of error/similarity measure for recognition. This module used the aligned data and model along with 3D kd- tree range search algorithm for computation of similarity between data and model.
Face Recognition on Low Quality Surveillance Images by Compensating Degradation
Face images obtained by an outdoor surveillance camera, are often confronted with severe degradations (e.g., low-resolution, low-contrast, blur and noise). This significantly limits the performance of face recognition (FR) systems used for binding “security with surveillance” applications. This work presents a framework to overcome the degradation for such images obtained by an outdoor surveillance camera, to improve the performance of FR. Two measures are defined based on the difference in intensity histograms and entropies of the gallery (large resolution, good contrast samples) and probe (with very low resolution, poor contrast and blur) images, to estimate the amount of degradation. Super-resolution techniques which are used to increase the image resolution of face samples, fail in these situations due to large difference (ratio) of the resolutions. We hence propose a combination of partial restoration (using super-resolution) of probe samples and degradation of gallery, to provide superior performance in FR. PCA (Principal component analysis) and FLDA (Fisher Linear Discriminant Analysis) have been used as baseline face recognition classifiers. The aim is to illustrate the effectiveness of our proposed method of compensating the degradation in surveillance data, rather than designing a specific classifier space suited for degraded test probes. The efficiency of the method is shown by an enhancement of the face classification accuracy, while comparing results obtained separately using training with acquired indoor gallery samples and then testing with the outdoor probes.
Top-k query processing with multi-dimensional range search
An m-dimensional top-k query (with m search conditions) is primarily processed by scanning the corresponding m index lists in descending score orders in an interleaved manner (and by making judicious random accesses to look up index entries of specific data items). In this paper a new algorithm is proposed that makes use of a data structure that facilitates multidimensional range search. An m-dimensional top-k query can be processed by searching for the data items that satisfies a range of score over each dimension. At every step of the algorithm a new set of ranges (one for each dimension) is specified such that more accurate tuples are added in the candidate top-k set. The process continues till we get the actual top-k data items. The incremented range set is specified with the help of the statistics of the distribution of data items in m-dimensional space. Thus, efficiency of the algorithm very much depends on the proper study and analysis of the distribution of the data items in the m-dimensional space and the data structure used.