Information Technology Reference
In-Depth Information
to illumination variations, there are claims that 3D face recognition is illumina-
tion invariant. Although 3D faces are illumination invariant once the data has
been acquired, the data acquisition process itself is not illumination invariant.
This is because accurate 3D face data requires active illumination from a laser
or a projector. Moreover, changes in ambient illumination can still have a great
impact on the accuracy and completeness of 3D data. Dark regions such as eye-
brows and specularities can cause missing data or spikes. These problems are
discussed in detail by Bowyer et al. [5] in their survey of 3D face recognition.
In search of illumination invariance, Chu et al. [7] proposed active frontal illu-
mination from NIR LEDs for face recognition. This approach has the advantage
of being invariant to ambient lighting and the NIR illumination is impercep-
tible to the eye. However, like 3D face recognition, this approach is not truly
illumination invariant as it relies on active illumination and custom hardware.
The human visual perception has inspired many researchers to use video or
image sequences to construct a joint representation of the face in spatial and tem-
poral space for identification [26]. A single image contains spatial information but
the temporal dimension defines trajectories of facial features and body motion
characteristics which may further assist classification. Arandjelovic and Cipolla
[1] proposed shape-illumination manifolds to represent a face under changing
illumination conditions. They first find the best match to a video sequence in
terms of pose and then re-illuminate them based on the manifold. Appearance
manifolds under changing pose were also used by Lee and Kriegman [12] to per-
form face recognition. Both approaches assume the presence of pose variations
which imply image acquisition over longer durations.
Li et al. [14] extracted the shape and pose free facial texture patterns from
multi-view face images and used KDA for classification. Liu et al. [16] per-
form online learning for multiple image based face recognition without using a
pre-trained model. Tangelder and Schouten [22] used a sparse representation of
multiple still images for face recognition. A common aspect of existing multiple
image/video-based techniques is that they rely on changes in pose or long term
changes to extract additional information which implies longer acquisition times.
An underlying assumption is that the images must contain non-redundant infor-
mation either due to the relative motion of the camera and the face or the motion
of the facial features due to expressions. Multiple images of a face acquired in-
stantly e.g. 10 frames/sec, from a fixed viewpoint, will be mostly redundant and
the temporal dimension will not contain any additional information.
It is possible to instantly acquire non-redundant images by changing the illu-
mination. Belhumeur and Kriegman used multiple images under arbitrary point
source illuminations to construct the 3D shape of objects [4]. Lee et al. [9] ex-
tended the idea to construct 3D faces and its corresponding albedo and subse-
quently used them to synthesize a large number (80-120) of facial images under
novel illuminations. The synthetic images were used to estimate the illumina-
tion cone of the face for illumination invariant face recognition. Hallinan [10]
empirically showed that the illumination cone can be approximated by a five
dimensional subspace. Basri and Jacobs [2] showed that the illumination cone of
Search WWH ::
Custom Search