Unknown Object Detection via Universal Scoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computer vision systems are limited in recognizing objects from unknown classes, as they rely on labeled images and specific detectors for known classes, failing to accurately identify objects outside their trained categories.
Innovation Solution
A system that uses known-object detectors and an unknown-object detector to generate scores for image regions, determining the likelihood of a complete object and unknown objects by comparing scores, and assigns labels based on scoring functions, allowing for the creation of new classes and retrieval of labels from external knowledge sources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If computer vision systems use labeled images and specific detectors for known classes, then they can accurately identify objects within trained categories, but they fail to recognize objects from unknown classes
Solution Approach 1:
The system applies a universal objectness detector to all detected regions regardless of their class label. This detector evaluates whether a region contains a complete object based on features like boundary completeness and contextual consistency, enabling the system to handle both known and unknown objects with a single mechanism rather than requiring separate detectors for each class.
Solution Approach 2:
The patent introduces an intermediary unknown-object detector that acts as a mediator between known-object detectors and the final classification system. This detector specifically identifies regions that may contain unknown objects by analyzing characteristics that differ from known classes, such as atypical feature patterns or inconsistencies with training data, thereby bridging the gap between structured classification and open-world recognition.
2Adaptability or versatility
If the system expands to detect unknown objects, then adaptability improves, but system complexity increases due to multiple detectors and scoring functions
Solution Approach 1:
The system merges the detection processes by having both known-object detectors and the unknown-object detector operate on the same image regions simultaneously. The results from multiple detectors are combined through a unified scoring mechanism that integrates class-specific scores with objectness scores, reducing redundant processing steps and simplifying the overall detection pipeline.
Solution Approach 2:
The system applies partial action by selectively applying the unknown-object detector only to regions that show ambiguous or low-confidence characteristics from known-object detectors. This approach avoids running all detectors on every region, thereby reducing computational overhead while maintaining the capability to detect unknown objects when necessary.
3Reliability
If the system uses multiple detectors and scoring functions, then detection capability improves, but processing time increases
Solution Approach 1:
The system performs preliminary filtering by having known-object detectors quickly assess regions and identify those with high confidence matches. Only regions that show low confidence or ambiguous characteristics are then subjected to the more computationally intensive unknown-object detector, thereby reducing the overall processing time while maintaining high detection reliability for both known and unknown objects.
Solution Approach 2:
The detection system operates continuously by processing regions in parallel and maintaining a unified scoring framework that continuously evaluates all regions. This continuous processing approach eliminates idle time between detection stages and ensures that the system can efficiently handle streaming or batch image data without repeated start-stop cycles.
Data Source
AI summary
Devices, systems, and methods obtain a region of an image; generate known-object scores for the region using known-object detectors, wherein each known-object detector of the known-object detectors detects objects in a respective object class; determine a likelihood that the region includes a complete object; and determine a likelihood that the region includes an unknown object based on the likelihood that the region includes a complete object and on the known-object scores.