Hierarchical ML Network for Rare Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computer vision systems struggle with accurately detecting and classifying rare objects and behaviors, leading to incorrect classifications or failure in object detection, especially when training data is sparse.
Innovation Solution
A machine-learning model architecture that employs a first ML model for general classification, followed by a sub-class ML model for refined classification, using feature maps and region of interest (ROI) inputs to improve detection accuracy and confidence, particularly for rare objects, by selecting appropriate sub-class models based on initial classifications and adjusting training through backpropagation of loss.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single ML model is used for general classification, then the system complexity is low, but the detection accuracy for rare objects is insufficient
Solution Approach 1:
The patent divides the classification task into multiple hierarchical levels: a first ML model performs general classification, and a second ML model performs sub-classification for rare objects. This segmentation allows each model to specialize in specific tasks, improving overall detection accuracy while managing complexity through functional division.
Solution Approach 2:
The first ML model acts as an intermediary that processes sensor data and identifies rare objects, then passes selected feature maps to the second ML model for further sub-classification. This intermediary approach enables the system to handle rare objects with specialized processing without requiring the entire system to be overly complex.
2Measurement precision
If more training data is collected to improve rare object detection, then the detection accuracy improves, but the time and resources required for data collection increase
Solution Approach 1:
The system performs preliminary classification with the first ML model to identify rare objects before applying the more computationally intensive second ML model. This preliminary action filters out common objects that don't require sub-classification, reducing overall training and inference time while maintaining high accuracy for rare objects.
Solution Approach 2:
The patent applies the second ML model selectively only to regions containing rare objects identified by the first model, rather than processing the entire image. This partial action approach achieves high detection accuracy for rare objects without the computational overhead of applying complex models to all image regions.
3Adaptability or versatility
If the ML model is trained to recognize all possible object variations, then the classification coverage is comprehensive, but the model becomes difficult to train with sparse data
Solution Approach 1:
The classification problem is segmented into two levels: general classification covering all object types and sub-classification covering specific variations of rare objects. This segmentation allows the first model to learn broad patterns from abundant data while the second model learns specific variations from sparse data, reducing overall training complexity.
Solution Approach 2:
The system applies different levels of classification detail to different object types: common objects receive general classification while rare objects receive detailed sub-classification. This local quality approach ensures comprehensive classification coverage where needed without unnecessarily complicating the model structure for all objects.
Data Source
AI summary
The techniques discussed herein may comprise refining a classification of an object detected as being represented in sensor data. For example, refining the classification may comprise determining a sub-classification of the object.