Steven C.H. Hoi, Wei Liu, Michael R. Lyu, WeiYing Ma

1
The Chinese University of Hong Kong
Learning Distance Metrics with Contextual
Constraints
for Image Retrieval
Steven C.H. Hoi, Wei Liu, Michael R. Lyu,
Wei-Ying Ma The Chinese University of Hong
KongMicrosoft Research Asia
Motivations
Contributions

• Distance metrics are important for image
  retrieval.
• Learning distance metrics with pairwise
  contextual constraints is critical to bridge
  the semantic gaps of image retrieval.
• Traditional distance metric learning usually
  studies linear distance metrics, which may
  not be effective for image retrieval.

• A new distance metric learning method is
  proposed for image retrieval.
• We developed two algorithms, Discriminative
  Component Analysis (DCA) and Kernel DCA, for
  learning metrics with contextual constraints.
• Empirical evaluations have been conducted for
  image retrieval.

.
METHODOLOGY
DCA

• Main Ideas
• Improving Relevant Component Analysis (RCA) by
  combining the dissimilar contextual constraints.
• Looking for the most discriminative
  transformation for learning the metrics.

Covariance matrix between data chunks
Covariance matrix within data chunks
KERNEL DCA
Learning the optimal transformation
Experimental Results
(a) Original Data Space
(a) Dogs retrieval
(b) Butterfly retrieval
(c) Roses retrieval
(b) Space via Kernel
Experimental results on the 20-Cat dataset
(c) Embedding Space via KDCA
The proposed DCA and Kernel DCA are promising for
learning distance metrics from contextual
constraints for image retrieval.
CoNCLUSION
CUHK and Microsoft Research Asia
IEEE Computer Vision and Pattern Recognition 2006