Approach

Visual Model Acquisition : Models are collections of images with similar feature distributions.

Model:
• Group images from a training sequence into groups.
• Represent each group with prototype distributions.
Filters:
• Low-confidence groups.
• non-discriminating groups.
Recognize:
• Register image with models.
• Match image with models using distance between distributions.

Motivation

Landmark and Environment Recognition in Outdoor Scenes:
• Localization : Compute vehicle location from landmarks.
• Mapping : Build map by collecting models of distinctive landmarks.
• Change Detection : Compare appearance of observed landmark with model.
• Cueing : Identify subsets of images from raw training set.
Issues:
• No a-priori knowledge of the discriminating landmarks.
• No a-priori geometric models.
• Considerable variation in appearance of objects

Feature Distributions

Feature Distributions for 16 subimages

Segments:
density, length, orientation, angles (18 dim for each)
Parallel segments
density, length, orientation (18 dim for each)
Contours
density, length, orientation (18 dim for each)
Color
histogram (18 dim for each)

Feature Distributions : Typical Examples

Original image





Color image after histogram quantization and normalization





Edge Image

• Compute mutual distances of feature distributions.
• Use relative weights between features.
• Use sequential constraints to limit search.

Images recognized under large variation in viewpoint and illumination: