Object Detection Device Dual Class Definition Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In object detection systems using machine learning, the performance of identifying specific classes and generalization is compromised when detecting a large number of classes, leading to over-learning and erroneous detections, such as mistaking a 'train' for a 'bus', due to the need for extensive training data and processing capacity allocation.
Innovation Solution
The system employs a dual class definition approach, where a first class definition is used for position area detection with a large number of classes and a second, smaller class definition for class identification, allowing for improved generalization and identification performance by focusing on specific classes and grouping others into a single category.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the class identification unit performs learning for identifying a large number of classes (e.g., 80 classes), then the detection capability covers more object types, but the identification performance for specific target classes deteriorates due to allocated performance being distributed across all classes
Solution Approach 1:
The patent segments the class identification process into two stages: first, an object position area detection unit detects objects across a large number of classes (e.g., 80 classes); second, a class identification unit identifies specific target classes (e.g., 6 classes) only for detected objects. This segmentation allows the system to maintain broad detection capability while concentrating identification performance on specific target classes, resolving the contradiction between coverage and precision.
2Measurement precision
If the object position area detection unit and class identification unit learn only from limited class data (e.g., six classes), then the identification performance for target classes is maintained, but over-learning occurs and generalization performance deteriorates
Solution Approach 1:
The patent applies preliminary action by having the object position area detection unit perform initial detection across a large number of classes (e.g., 80 classes) before the class identification unit performs specific class identification. This preliminary detection with diverse class data prevents over-learning in the subsequent specific class identification, as the system has already encountered various object types in the first stage, thereby maintaining both specific class performance and generalization capability.
3Use of energy by moving object
If the scale (processing capacity) of the class identification unit is kept constant while learning for many classes, then resource consumption is controlled, but identification performance is diluted across all learned classes
Solution Approach 1:
The patent segments the class handling into two distinct units with different scopes: the object position area detection unit handles a large number of classes (e.g., 80 classes) with broader processing, while the class identification unit focuses on a small number of target classes (e.g., 6 classes) with concentrated processing capacity. This segmentation allows the class identification unit to maintain constant scale and high identification performance for target classes without needing to allocate resources for learning all 80 classes, thus controlling resource consumption while preserving precision.
Data Source
AI summary
An object position area detection unit of an object detection device detects a position area of an object included in an inputted image, on the basis of a first class definition in which a plurality of classes are defined in advance. A class identification unit identifies which of the plurality of classes the object belongs to, on the basis of a second class definition in which a plurality of classes are defined in advance. An object detection result output unit outputs an object detection result on the basis of a detection result of the object position area detection unit and an identification result of the class identification unit. The number of classes defined by the second class definition is smaller than the number of classes defined by the first class definition. The plurality of classes defined by the second class definition are formed by collecting some of a plurality of classes defined by the first class definition.