Simon Baker

We propose a new algorithm for recognition of low-resolution faces for cases when multiple images of the same face are available at matching. Specifically, this is the case of multiple-frame, or video, face recognition, and recognition with multiple cameras. Face recognition degrades when probe faces are of lower resolution than those available for training. There are two paradigms to alleviate this problem, but both have clear disadvantages. One is to use super-resolution algorithms to enhance the image, but as resolution decreases, super-resolution becomes more vulnerable to environmental variations, and it introduces distortions that affect recognition. On the other hand, it is possible to match in the low-resolution domain by downsampling the training set, but this is undesirable because features important for recognition depend on high frequency details that are erased by downsampling. We recently proposed a new framework for recognition that is different from these two paradigms, and we have shown that recognition is considerably improved for still-image face recognition. In this work, we show that the proposed method can be generalized to use a stream of frames, and produce even better recognition performance. The new algorithm incorporates the underlying assumptions of super-resolution methods with subspace distance metrics used for classification, simultaneously. We extend our previous formulation to use multiple frames, and we show that it can also be generalized to use multiple image formation processes, modeling different cameras. In this manuscript we consider two scenarios: face recognition with multiple frames, and the case of having a camera sensor network with dual cameras in a master-slave setting. Using the Multi-PIE database, our results show an increase of rank-1 recognition accuracy of about 6% for both cases compared to the single frame settings.

Active Appearance Models (AAMs) are generative, parametric models that have been successfully used in the past to model deformable objects such as human faces. The original AAMs formulation was 2D, but they have recently been extended to include a 3D shape model. A variety of single-view algorithms exist for fitting and constructing 3D AAMs but one area that has not been studied is multi-view algorithms. In this paper we present multi-view algorithms for both fitting and constructing 3D AAMs. Fitting an AAM to an image consists of minimizing the error between the input image and the closest model instance; i.e. solving a nonlinear optimization problem. In the first part of the paper we describe an algorithm for fitting a single AAM to multiple images, captured simultaneously by cameras with arbitrary locations, rotations, and response functions. This algorithm uses the scaled orthographic imaging model used by previous authors, and in the process of fitting computes, or calibrates, the scaled orthographic camera matrices. In the second part of the paper we describe an extension of this algorithm to calibrate weak perspective (or full perspective) camera models for each of the cameras. In essence, we use the human face as a (non-rigid) calibration grid. We demonstrate that the performance of this algorithm is roughly comparable to a standard algorithm using a calibration grid. In the third part of the paper, we show how camera calibration improves the performance of AAM fitting. A variety of non-rigid structure-from-motion algorithms, both single-view and multi-view, have been proposed that can be used to construct the corresponding 3D non-rigid shape models of a 2D AAM. In the final part of the paper, we show that constructing a 3D face model using non-rigid structure-from-motion suffers from the Bas-Relief ambiguity and may result in a “scaled” (stretched/compressed) model. We outline a robust non-rigid motion-stereo algorithm for calibrated multi-view 3D AAM construction and show how using calibrated multi-view motion-stereo can eliminate the Bas-Relief ambiguity and yield face models with higher 3D fidelity.

Acquiring 3D models of intricate objects (like tree branches, bicycles and insects) is a hard problem due to severe self-occlusions, repeated thin structures and surface discontinuities. In theory, a shape-from-silhouettes (SFS) approach can overcome these difficulties and use many views to reconstruct visual hulls that are close to the actual shapes. In practice, however, SFS is highly sensitive to errors in silhouette contours and the calibration of the imaging system, and therefore not suitable for obtaining reliable shapes with a large number of views. We present a practical approach to SFS using a novel technique called coplanar shadowgram imaging, that allows us to use dozens to even hundreds of views for visual hull reconstruction. Here, a point light source is moved around an object and the shadows (silhouettes) cast onto a single background plane are observed. We characterize this imaging system in terms of image projection, reconstruction ambiguity, epipolar geometry, and shape and source recovery. The coplanarity of the shadowgrams yields novel geometric properties that are not possible in traditional multi-view camera- based imaging systems. These properties allow us to derive a robust and automatic algorithm to recover the visual hull of an object and the 3D positions of light source simultaneously, regardless of the complexity of the object. We demonstrate the acquisition of several intricate shapes with severe occlusions and thin structures, using 50 to 120 views.

We introduce a semi-supervised method for building large, labeled datasets effaces by leveraging archival video. Specifically, we have implemented a system for labeling 11 years worth of archival footage from a television show. We have compiled a dataset of 611,770 faces, orders of magnitude larger than existing collections. It includes variation in appearance due to age, weight gain, changes in hairstyles, and other factors difficult to observe in smaller-scale collections. Face recognition in an uncontrolled setting can be difficult. We argue (and demonstrate) that there is much structure at varying timescales in the video data that make recognition much easier. At local time scales, one can use motion and tracking to group face images together - we may not know the identity, but we know a single label applies to all faces in a track. At medium time scales (say, within a scene), one can use appearance features such as hair and clothing to group tracks across shot boundaries. However, at longer timescales (say, across episodes), one can no longer use clothing as a cue. This suggests that one needs to carefully encode representations of appearance, depending on the timescale at which one intends to match. We assemble our final dataset by classifying groups of tracks in a nearest-neighbors framework. We use a face library obtained by labeling track clusters in a reference episode. We show that this classification is significantly easier when exploiting the hierarchical structure naturally present in the video sequences. From a data-collection point of view, tracking is vital because it adds non-frontal poses to our face collection. This is important because we know of no other method for collecting images of non-frontal faces "in the wild".