Research Topics 


  • Face recognition with a single model image
    (Single-to-Single Matching) :

The recognition task with a model image has received increasing attention for the important applications such as automatic passport control at airport, where a single photo in the passport is available as a model, and face image retrieval in Internet/or unknown database. In the retrieval task, a single arbitrary query image is supplied by users and every single image in the unknown database is matched with the single query, bringing a Single-to-Single matching problem. The task has emerged as an active research area in FRT (Face Recognition Test) and MPEG-7 (Moving Picture Experts Group) Standardisation for face image retrieval. The example patterns are shown in Figure 1.

Selected Publication: T-K. Kim and J. Kittler, Locally Linear Discriminant Analysis for Multi-modally Distributed Classes for Face Recognition with a Single Model Image, IEEE Trans. on PAMI, 27(3):318-327, 2005 (Regular paper).

(a) Single-to-Single matching strategy

(b) Single face images captured at different views (XM2VTS).

Figure 1. Example Patterns For Single-to-Single Matching. Given a single model image per a subject (for e.g. a frontal view), novel view images should be classified.


  • Object recognition by image sets (Set-to-Set Matching) :

Rather than a single image input, more robust object recognition can be achieved classifying a set of images which represents a variation in objectís appearance. Sets may be derived from sparse and unordered observations acquired by multiple still shots as well as from a video sequence. That is, the objective of this task is to classify an unknown set of images to one of the training classes, each of which is also represented by images sets. The example sets of an isolated general object are shown in Figure 2.

Selected Publication: T-K. Kim, J. Kittler and R. Cipolla, Discriminative Learning and Recognition of Image Set Classes Using Canonical Correlations, IEEE Trans. on PAMI, 29(6), 2007 (Regular paper).

(a) Set-to-Set matching strategy

(b) A set of object images collected from a turn-table sequence.

(c) A set of object images collected from a random-moving-camera sequence.

Figure 2. Examples Patterns For Set-to-Set Matching. The two sets contain different pattern variations caused by different views and lighting.


  • Action classification in videos (Video-to-Video Matching) :

Over the last decades, human action/gesture classification has been one of the most important topic in computer vision for a variety of tasks such as video surveillance, object level video summarization, video indexing, digital library organization, etc. This task can be tackled by classifying spatiotemporal patterns in aligned videos. Action detection may be first performed out to localize unit actions in input videos. Example video samples with indication of action alignments are shown in Figure 3 (Also see our own hand-gesture data set).

Selected Publication: T-K. Kim, S-F. Wong and R. Cipolla, Tensor Canonical Correlation Analysis for Action Classification, In Proc. of IEEE Conference on CVPR,  Minneapolis, MN, 2007. 

S-F. Wong, T-K. Kim and R. Cipolla, Learning Motion Categories Using Both Semantic and Structural Information, In Proc. of IEEE Conference on CVPR, Minneapolis, MN, 2007.

(a) Video-to-Video matching strategy

(b) Example action classes (KTH). The bounding box and superimposition indicate the action alignment in a spatiotemporal space.

Figure 3. Examples Patterns For Video-to-Video Matching.