I have been recently working with Jon on a latent topic model for simultaneous object detection and segmentation. We are studying the use of a certain type of Hierarchical Bayesian Model (a variant of Blei's Latent Dirichlet Allocation) with dense image features (to be described in the near future) and applying our algorithm to the 2006 PASCAL Visual Object Classes Challenge.
Since I've been introduced to DDMCMC for image segmentation, I have been thinking about the relationship between segmentation and recognition. Also, in my Advanced Perception course, we recently looked at some Borenstein/Ullman papers that incorporate segmentation and recognition into one framework (this resulted in a few more ideas). Here are some short ideas about segmentation:
How is the problem of segmentation usually posed? Given an image, produce a partition of the image into K disjoint regions. If one wants to only use the image data given with no object-level assumptions, then one can only proceed to find the maximum likelihood segmentation. However, such a segmentation will have a very large variance (and a small bias) because there are many different ways of 'grouping' local image structures together.
Fortunately, one can be Bayesian and relate the 'seemingly independent' problems of segmenting different images by utilizing a prior over image regions. When employing a Bayesian Hierarchical model, one generally breaks down the problem into two stages: parameter estimation(training) and statistical inference(testing). Just like in the LDA model, parameter estimation is concerned with estimating the hyperparameters of the hierarchical model (and also finding the distributions over latent variables for each document in the training set) and inference is segmenting a novel image (by utilizing the parameters obtained in the training stage). By being Bayesian, one will introduce bias when segmenting a novel image (the bias will make the novel segmentation more like some of the segmentations that were obtained for the training corpus) and reduce variance. Isn't bias a good thing in this case? Don't we REALLY want a novel segmentation to be somehow related to other segmentations? Perhaps the only property about a segmentation engine that we care about is its object-level consistency across a wide number of images. We hope that: given a large enough training corpus that captures large variability in pose and appearance for a large number of objects, the 'semantic'-segmentation that is desired will be the one that is approximated with our hierarchical model. A latent topic-based segmentation engine would additionaly provide a registration across image features via the latent topic space. In other words, semantically equivalent image primitives would be near each other in some high dimensional latent-topic parameter space.
The new question should be: How can we learn to segment novel images given a corpus of images that are somehow related? Under this view, a segmentation of a novel image is a k-way parition of an image AND the latent topic distributions associated with each segment (registration of segments across images). Remember, object tracking is temporal registration and object recognition is semantic registration.