Sandra Mau

Patch-based probabilistic image quality assessment for face selection and improved video-based face recognition

In video based face recognition, face images are typically captured over multiple frames in uncontrolled conditions, where head pose, illumination, shadowing, motion blur and focus change over the sequence. Additionally, inaccuracies in face localisation can also introduce scale and alignment variations. Using all face images, including images of poor quality, can actually degrade face recognition performance. While one solution it to use only the ‘best’ of images, current face selection techniques are incapable of simultaneously handling all of the abovementioned issues. We propose an efficient patch-based face image quality assessment algorithm which quantifies the similarity of a face image to a probabilistic face model, representing an ‘ideal’ face. Image characteristics that affect recognition are taken into account, including variations in geometric alignment (shift, rotation and scale), sharpness, head pose and cast shadows. Experiments on FERET and PIE datasets show that the proposed algorithm is able to identify images which are simultaneously the most frontal, aligned, sharp and well illuminated. Further experiments on a new video surveillance dataset (termed ChokePoint) show that the proposed method provides better face subsets than existing face selection techniques, leading to significant improvements in recognition accuracy.

Face recognition from still images to video sequences: a local-feature-based framework

Although automatic faces recognition has shown success for high-quality images under controlled conditions, for video-based recognition it is hard to attain similar levels of performance. We describe in this paper recent advances in a project being undertaken to trial and develop advanced surveillance systems for public safety. In this paper, we propose a local facial feature based framework for both still image and video-based face recognition. The evaluation is performed on a still image dataset LFW and a video sequence dataset MOBIO to compare 4 methods for operation on feature: feature averaging (Avg-Feature), Mutual Subspace Method (MSM), Manifold to Manifold Distance (MMS), and Affine Hull Method (AHM), and 4 methods for operation on distance on 3 different features. The experimental results show that Multi-region Histogram (MRH) feature is more discriminative for face recognition compared to Local Binary Patterns (LBP) and raw pixel intensity. Under the limitation on a small number of images available per person, feature averaging is more reliable than MSM, MMD, and AHM and is much faster. Thus, our proposed framework—averaging MRH feature is more suitable for CCTV surveillance systems with constraints on the number of images and the speed of processing.

Video face matching using subset selection and clustering of probabilistic Multi-Region Histograms

Balancing computational efficiency with recognition accuracy is one of the major challenges in real-world video-based face recognition. A significant design decision for any such system is whether to process and use all possible faces detected over the video frames, or whether to select only a few ‘best’ faces. This paper presents a video face recognition system based on probabilistic Multi-Region Histograms to characterise performance trade-offs in: (i) selecting a subset of faces compared to using all faces, and (ii) combining information from all faces via clustering. Three face selection metrics are evaluated for choosing a subset: face detection confidence, random subset, and sequential selection. Experiments on the recently introduced MOBIO dataset indicate that the usage of all faces through clustering always outperformed selecting only a subset of faces. The experiments also show that the face selection metric based on face detection confidence generally provides better recognition performance than random or sequential sampling. Moreover, the optimal number of faces varies drastically across selection metric and subsets of MOBIO. Given the trade-offs between computational effort, recognition accuracy and robustness, it is recommended that face feature clustering would be most advantageous in batch processing (particularly for video-based watchlists), whereas face selection methods should be limited to applications with significant computational restrictions.

Invited paper: Embedded face and biometric technologies for national and border security

The CCTV surveillance industry is undergoing a sea change due to the adoption of IP technologies. This is allowing the integration of a plethora of new cameras and other sensors into huge integrated networks. Adoption of IP technologies is presenting opportunities for scalable visual analytics that has the potential to add enormous value to entire camera networks. One such technology is scalable robust face search to identify persons of interest in large crowds. Not only are such systems required to work robustly in a wide variety of conditions, they must also be extremely fast and scalable to hundreds, if not thousands, of high definition camera nodes. Developing and testing such technology is challenging and requires a combination of fast algorithms, distributed databases, mobile platform integration, parallel processing using distributed middleware such as ROS, and GPU acceleration using tools such as CUDA and OpenCL. In this paper we cover emerging system trends such as super-megapixel cameras, post incident digital PTZ, integration and fusion of video and non-video sensors, multimodal remote biometrics including face and iris on the move. Finally, recognition results will be presented from a formal face recognition trial in early 2011 in one of Asias largest International airports.

A face biometric benchmarking review and characterisation

In order to advance face recognition research, algorithm performance has to be measured and compared using a range of metrics and operating characteristics. While public challenges such as the NIST-sponsored FERET, FRVT, FRGC, and MBGC are helpful to gauge comparative performance and improvement for a particular scenario, they typically are not sufficient to fully characterise the strengths and weaknesses of the face recognition algorithm, thus researchers need to do additionally benchmarking independently. This paper provides: (1) a detailed review and categorisation of publicly available face biometrics benchmarks; (2) a discussion of metrics and performance factors to consider; (3) a proposal for a meta-face biometric benchmarking regime which suggests guidelines for benchmarking across multiple datasets to more fully characterise and quantify face recognition performance across various operating characteristics; and (4) a sample demonstration which compare the performance of a face recognition algorithm before and after inclusion of a face quality metric.

Gaussian Probabilistic Confidence Score for Biometric Applications

We propose a quick and widely applicable approach for converting biometric identification match scores to probabilistic confidence scores, resulting in increased discrimination accuracy. This approach builds on a confidence scoring approach for Binomial distributions resulting from Hamming distances (commonly used in iris recognition). We derive a Gaussian confidence scoring approach that is three orders of magnitude faster than the Binomial approach while still resulting in higher recognition rates. Gaussian distributions are also more common and thus more widely applicable to different biometric systems. For probe-to-gallery (1-to-N) identification of the face recognition system tested, this approach has been shown to improve the identification rate from 25.66% to 68.05% at 1.00% false alarm rate for a CCTV video matching dataset, and from 63.34% to 73.14% for images from the LFW dataset. A sensitivity analysis demonstrates that modeling errors in genuine and impostor distributions only negatively impacts discrimination when the distribution means are modelled to be closer together than the true underlying distributions. For the reverse case where the distribution means are modeled to be further apart than the true distributions, discrimination accuracy is improved.

A framework for lab-based real-time video analysis on distributed camera networks

In the field of video analytics for surveillance, the trend towards the use of multi-camera and high definition video is increasing. This poses significant technical challenges in terms of video transmission and real-time processing for surveillance analytics, such as people recognition and tracking. Currently, available solutions are typically proprietary commercial systems which are costly to purchase. These proprietary systems also do not facilitate research collaboration across members of the computer vision community. We propose a framework for video analytics research based only on open-source software which is collaborative, scalable, interoperable, and distributed. This framework was successfully applied to the task of face recognition on both live video feeds and video datasets.

Improved Person Re-Identification Using Statistical Approximation

Person re-identification on image sets in which each image is taken from a different angle and lighting condition is a very challenging task. This task becomes even more difficult when images are low resolution and carrying image compression artifacts. The accuracy of the existing re- identification techniques are relatively low on the challenging evaluation grounds such as VIPeR and iLIDS image datasets. In these datasets, distortions in shape and colour make the re- identification task difficult and uncertain for both machine and human. In this paper, we propose a new approach to address the uncertainty in low resolution images for person re-identification by using statistical approximation. We first show that the distribution within a patch on persons image does not fit a normal distribution via Kolmogorov- Smirnov test. Then we simplify the Kolmogorov- Smirnov statistic by using only the mean and standard deviation of the distribution. These values are used as descriptors for per region per channel, and concatenated for comparison of image pairs. Experiments show that the proposed approach outperforms the state-of-the-art on person re- identification methods. The small memory foot print and the low computational cost of the proposed technique make it suitable for person re- identification in large scale surveillance applications.