Object Detector Training Head for Class Confusion Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network-based object detectors face challenges in correctly distinguishing similar object classes, leading to incorrect classifications, especially on limited hardware, and existing methods to address this issue often introduce resource inefficiencies or performance degradation.
Innovation Solution
A method involving a training head added to the neural network during training to determine the truthfulness of object class assignments, using modified labels and objective functions to reduce class confusion without altering the network's architecture or loss functions during inference.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If additional modules are added to the neural network to reduce class confusion, then classification accuracy is improved, but device complexity and resource consumption increase
Solution Approach 1:
The patent applies preliminary action by performing data augmentation and generating modified training data before the main training process. Synthetic examples of confusing object classes are created in advance and integrated into the training dataset, allowing the neural network to learn distinguishing features earlier in the training process without requiring additional architectural modules during inference
Solution Approach 2:
The patent uses copying by creating synthetic copies of training data through data augmentation techniques. Modified training data including synthesized confusing object examples are generated from existing data, enabling the network to learn from expanded examples without duplicating the entire network architecture
2Measurement precision
If visual attention modules are added to reduce class confusion, then classification accuracy is improved, but runtime increases
Solution Approach 1:
The patent performs data augmentation and synthetic example generation during the training phase before deployment. By preparing enhanced training data in advance with synthesized confusing class examples, the network learns attention to distinguishing features during training without requiring runtime attention mechanisms that would increase inference time
3Measurement precision
If loss functions are modified to focus on confusing classes, then class confusion is reduced, but detection quality may degrade
Solution Approach 1:
The patent segments the training data into different categories including confusing object classes and non-confusing classes. By separating and specifically augmenting confusing class examples, the training process can focus attention where needed without applying aggressive loss modifications that would harm overall detection quality across all object classes
Solution Approach 2:
The patent applies data augmentation selectively to confusing object classes rather than uniformly across all classes. This local approach enhances distinguishing features for problematic classes while maintaining natural training data characteristics for other classes, preventing over-correction that would degrade overall detection performance
4Measurement precision
If training data is augmented with synthesized confusing object classes, then class confusion is reduced, but training data volume increases
Solution Approach 1:
The patent applies partial action by selectively augmenting only the confusing object classes that benefit most from synthetic examples, rather than uniformly augmenting all training data. This targeted approach reduces the total volume of augmented data while still achieving the goal of reducing class confusion for problematic categories
Data Source
Figure 1
Figure 2
Figure 3a~3c
AI summary
The present invention is concerned with a computer-implemented method for training an object detector (OD), an object detector (OD), a computer program and a computer-readable (storage) medium. To carry out the method a set of object classes (C) confused by the neural network (NN), an object feature map (FM) from the neural network (NN), and an object class label (LO) comprising object classes (C) assigned to each object (O) in the object feature map (FM) are obtained. Subsequently, a training head (TH) configured to determine a truthfulness of an object class (C) assigned to an object (O) in an object feature map (FM) based on an object feature map (FM) and an object class label (LO) is added to the neural network (NN). Based on the set of object classes (C) and the object class label (LO) a modified object class label (LO') for which at least one object class assignment (C) is modified is determined. This modified object class label (LO') together with the object feature map (FM) is processes through the neural network generating an output which is compared with a truthfulness label (LT) comprising the indication of the truthfulness of each object class (C) of the modified object class label (LO') by using a first objective function (OF1). Network parameters are then updated based on the comparison.