Likelihood Map Segmentation for Object Tracking Correspondence
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing tracking techniques struggle to accurately associate specific parts of the same object when nearby objects are present, leading to erroneous correspondences.
Innovation Solution
A learning apparatus and method that utilize a first model for detecting specific parts and a second model for tracking targets, generating a likelihood map and region map, and performing learning based on an element product map to improve correspondence accuracy by filtering specific parts of the tracking subject.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a bottom-up approach is used to detect specific parts and associate them to determine the specific part of the subject, then the detection of specific parts can be achieved, but the specific parts may be erroneously associated with a specific part of an object different from the subject
Solution Approach 1:
The patent segments the correspondence determination process into two independent models: a first model that detects specific parts and generates a likelihood map, and a second model that determines correspondence between specific parts and subjects. This segmentation allows each model to specialize in its respective function, improving overall accuracy while preventing erroneous associations through the independent verification provided by the second model.
Solution Approach 2:
The patent introduces a likelihood map as an intermediary between specific part detection and correspondence determination. The likelihood map represents the probability of each detected specific part belonging to the subject, serving as a mediator that provides probabilistic evidence for correspondence decisions. This intermediary enables more reliable association by quantifying the likelihood of correct correspondence rather than making direct associations.
2Productivity
If multiple specific parts are detected separately and associated with each other, then the specific parts can be identified, but the specific parts may be erroneously corresponded with a part of another object present nearby
Solution Approach 1:
The patent divides the correspondence determination into separate specialized models: the first model rapidly detects multiple specific parts and generates a likelihood map, while the second model independently determines correspondence. This segmentation maintains high detection speed through parallel processing while improving correspondence accuracy through the specialized second model that focuses solely on association decisions.
Solution Approach 2:
The patent changes the parameter representation from direct binary association to probabilistic likelihood values. The likelihood map stores probability values for each specific part belonging to the subject, allowing the system to make more informed correspondence decisions by considering the degree of likelihood rather than making definitive assignments, thereby reducing erroneous correspondences with nearby objects.
Data Source
AI summary
A learning apparatus comprises one or more memories storing instructions and one or more processors that execute the instructions to acquire a likelihood map of a specific part in an input image by using a first model for detecting the specific part, acquire a region map representing a region of a specific part of a tracking target in the input image by using a second model for detecting the tracking target, and perform learning of the second model based on a loss obtained based on an element product map obtained by an element product of the likelihood map and the region map and correct answer data indicating a region of a specific part of a tracking target in the input image.