Unsupervised Detection and Tracking of Arbitrary Objects with Dependent Dirichlet Process Mixtures

TL;DR

This paper introduces an unsupervised, general framework for detecting and tracking arbitrary objects in videos using dependent Dirichlet process mixtures, eliminating the need for object-specific detection methods.

Contribution

It presents a novel application of DDPMs for unsupervised object detection and tracking across diverse video scenarios, demonstrating broad applicability without modification.

Findings

Successfully detects and tracks diverse objects in synthetic and benchmark videos

Operates effectively over non-uniform backgrounds and occlusions

Requires no object-specific tuning or prior training

Abstract

This paper proposes a technique for the unsupervised detection and tracking of arbitrary objects in videos. It is intended to reduce the need for detection and localization methods tailored to specific object types and serve as a general framework applicable to videos with varied objects, backgrounds, and image qualities. The technique uses a dependent Dirichlet process mixture (DDPM) known as the Generalized Polya Urn (GPUDDPM) to model image pixel data that can be easily and efficiently extracted from the regions in a video that represent objects. This paper describes a specific implementation of the model using spatial and color pixel data extracted via frame differencing and gives two algorithms for performing inference in the model to accomplish detection and tracking. This technique is demonstrated on multiple synthetic and benchmark video datasets that illustrate its ability to, without modification, detect and track objects with diverse physical characteristics moving over non-uniform backgrounds and through occlusion.