Bayesian Network Inference for Unknown Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object recognition technologies, such as semantic segmentation, struggle to recognize objects not registered in their dictionary data, leading to incomplete self-driving capabilities when unknown objects are encountered.
Innovation Solution
An information processor and mobile device configuration that uses a Bayesian network to infer unknown objects by creating a network linking detected objects and comparing it with learned reference networks, allowing for the identification of unknown objects through similarity selection and inference.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If semantic segmentation is used for object recognition, then objects in the image can be classified into categories, but objects that are not registered as dictionary data cannot be recognized
Solution Approach 1:
The patent introduces an intermediary inference process using Bayesian networks between the image input and final object recognition. When semantic segmentation fails to recognize an object (unknown object detected), the system creates a Bayesian network that models relationships between detected objects and infers the unknown object's identity through probabilistic reasoning, rather than direct classification. This intermediary mechanism enables recognition of objects not in the dictionary while maintaining accuracy for known objects.
2Adaptability or versatility
If a Bayesian network is used to perform inference for unknown data, then unknown objects can be recognized, but the processing complexity increases
Solution Approach 1:
The patent applies Bayesian network inference selectively rather than universally. The system first attempts semantic segmentation for all objects, and only when an unknown object is detected does it activate the Bayesian network inference process. This partial application of the complex inference mechanism reduces overall processing complexity while maintaining the ability to recognize unknown objects when necessary.
3Speed
If semantic segmentation is used for object recognition, then processing speed is maintained, but unknown objects cannot be recognized leading to safety issues in self-driving
Solution Approach 1:
The patent prepares Bayesian network models and reference data in advance during an offline learning phase. During real-time self-driving operation, the system only needs to perform comparison and inference using pre-processed models, rather than building complex probabilistic models from scratch. This preliminary preparation maintains processing speed during critical driving operations while enabling reliable recognition of unknown objects through the pre-trained inference mechanisms.
Data Source
AI summary
To provide a configuration that enables recognition of an unknown object included in an image. The configuration includes a network creation unit that creates a network linking a scene node indicating a scene type of an image and a node corresponding to an object, and an unknown object inference unit that compares a created network created by the network creation unit with a large number of reference networks created by learning processing, selects reference networks similar to the created network, selects, from the selected reference networks, a node corresponding to a node of an unknown object in the created network, and infers that an object associated with the selected node is an object corresponding to the unknown object.