Research Interests
I am interested in statistical modeling, pattern recognition, computer vision and cognitive neuroscience.
Highlights
|
This work proposes an active basis model, a shared sketch algorithm, and a computational architecture of sum-max maps for representing, learning, and recognizing deformable templates. |
|
In a unified information theoretic framework, we learn hybrid image templates composed of sketch and texture descriptors. |
|
Using this image representation we may discover visual objects under weak supervision, i.e. without knowing the true locations, orientations and categories of the visual objects. |
Generative models of images provide insights into the image representation, i.e. how images should be stored. Images, different from natural language, speech, gene expression etc., are more efficiently represented by certain features (e.g. 2D geometric primitives, interest regions, color) and certain composition rules binding the features. It is then interesting to spot them among thousands of candidate features and their compositions.
The Active basis model is a probablistic shape template learnable from training examples from an image category, into which we may add image features that model the texture ("geometry-less") part of images and thus forming a mixed template. The learning of both models follow information theoretic criteria. The image templates can also be learned under weak or no supervision via EM-type algorithms, to automatically discover a large numebr of image templates as visual concepts. For a huge collection of such learned structures, we need to find a way to efficiently rank them, index them, and query them. A stochastic image grammar is a powerful tool to achieve this.