08-16-2017, 10:45 PM
Unsupervised Learning of Probabilistic Object Models
[attachment=16243]
INTRODUCTION
RECENT work on object classification and recognition has
tended to represent objects in terms of spatial configurations
of features at a small number of interest points [3],
[4], [5], [6], [7], [8]. Such models are computationally efficient,
for both learning and inference, and can be very effective for
tasks such as classification. But they have two major
disadvantages: 1) the sparseness of their representations
restricts the set of visual tasks they can perform and 2) these
models only exploit a small set of image cues
2 LEARNING BY KNOWLEDGE PROPAGATION
We now describe our strategy for learning by knowledge
propagation. Suppose that our goal is to learn a generative
model to explain some complicated data. It may be too hard
to attempt a model that can explain all the data in one
attempt. An alternative strategy is to build the model
incrementally by first modeling those parts of the data,
which are the easiest. This will provide the context making
it easier to learn models for the rest of the data.
3 THE IMAGE REPRESENTATION
This section describes the different image features that we
use: 1) interest points (used in POM-IP), 2) regional features
(in POM-mask), and 3) edgelets (in POM-edgelet).
The interest point features d1 I of an image I used in POMIP
are represented by a set of attributed features d1 I fzig,
where zi xi; i;Ai with xi the position of the feature in the
image, i is the feature s orientation, and Ai is an appearance
vector. The procedures used to detect and represent the
feature points were described in [1], [2].
5 POM-MASK
The POM-mask uses regional cues to perform segmentation/
localization. It is trained using knowledge from the
POM-IP giving crude estimates for the segmentation (e.g.,
the bounding box of the IPs). This training enables POMmask
to learn a shape prior for each aspect of the object.
After training, the POM-mask and POM-IP are coupled
Fig. 4. During inference, the POM-IP supplies estimates of
pose and aspect to help estimate the POM-mask variables.