Computer Science ETDs
Publication Date
Summer 7-28-2026
Abstract
Advances in hardware and algorithms have produced an explosion in computer vision applications, with object recognition systems in operation across millions of platforms, ranging from handheld devices to satellite constellations. Many of these applications operate in an open-world context, where the algorithms come into contact with unknown objects. In areas such as medical imaging or object recognition, recognizing when something is not a known class present in the training data is important for reliable and safe operation. Many computer vision models and algorithms have no capability for the detection of unknown classes, and are known as closed-set classifiers. Out-of-distribution detection converts closed-set classifiers into open-set classifiers by providing the missing detection capability in the form of an altered vision model or a standalone module which understands the boundaries of the vision module's knowledge, and rejects unknown classes from being classified. Multinomial feature matching applies a statistical model derived from Dirichlet-Multinomial probability distributions to the learned features in convolutional and transformer networks, enabling feature-derived and probability-driven detection of out-of-distribution data. It is lightweight, easy to compute, and performs well when compared to similar out-of-distribution detection algorithms from prior work. It also offers explainability that most other detection schemes lack, and can optionally be used to classify in-distribution data as well, operating as a hybrid classification model with built-in out-of-distribution detection capability.
Language
English
Keywords
computer vision, machine learning, out-of-distribution detection, object recognition
Document Type
Dissertation
Degree Name
Computer Science
Level of Degree
Doctoral
Department Name
Department of Computer Science
First Committee Member (Chair)
Gruia-Catalin Roman
Second Committee Member
Trilce Estrada
Third Committee Member
Devin White
Fourth Committee Member
Shuang Luan
Fifth Committee Member
Manel Martinez-Ramon
Recommended Citation
Pitts, Christopher. "Multinomial Feature Matching for Out-of-Distribution Detection in Computer Vision Systems." (2026). https://digitalrepository.unm.edu/cs_etds/146