Image fragment matching in practice: applications and supplementary tools
Andrzej Sluzek
(Nanyang TU, Singpore)
Abstract:
Image fragment matching is a very general tool with a wide range of
prospective applications (sometimes very different from the originally
envisaged use in visual information retrieval). In this presentation, we
focus on three highly diversified applications. First, the idea of visual
object formation is discussed, where the system generalizes similarities
between fragments from dataset images to form "visual objects", i.e.
pieces of mutually similar fragments repetively encountered in the
dataset. Such "visual objects" are defined without any knowledge of image
domain/contents. In most cases, they correspond to the actual objects
frequently appearing in the dataset images, i.e. a certain level of
"visual understanding" of the dataset domain can be automatically
established. In the second application, we show that image fragment
matching can be a useful mechanism in face detection and/or recognition.
We argue that it can objectively model the subjective concepts of
"similar faces", "similar eyes", "similar foreheads", etc. Fragments of
such faces can be identified in images of unpredictable backgrounds.
Additionally, if certain assumptions are added, a very simple face
authentication algorithm can be built (of performances below the top
state-of-the-art results, but highly useful in diversified applications).
The third example presents how image fragment matching can be used to
estimate the visibility quality in traffic applications. Two numerical
measures of visibility are proposed, and the results on extremely
difficult data are presented. The approach can be prospectively used to
automatically adjust the vehicle speed to the current road conditions.
Finally, a few supplementary algorithms are briefly discussed. The
algorithms can be either used as alternatives in selected steps of image
fragment matching or can further improve the results of matching.