Object Detector Training Head for Class Confusion Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network-based object detectors face challenges in accurately distinguishing similar object classes, leading to incorrect classifications and potential failures in applications like advanced driver assistance systems, particularly on limited hardware.
Innovation Solution
A method is introduced that involves determining confused object classes, adding a training head to the neural network to assess the truthfulness of object class assignments, and using a combined objective function to update parameters, without modifying the network during inference.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If additional modules are added to the neural network to reduce class confusions, then classification accuracy is improved, but device complexity and resource consumption increase
Solution Approach 1:
The patent segments the object detector into two distinct parts: the original neural network for detection and a separate training head for confusion reduction. The training head is attached only during the training phase and can be removed afterward, allowing the main network to remain simple while still achieving improved classification accuracy through the specialized training mechanism.
Solution Approach 2:
The training head acts as an intermediary component that mediates between the original neural network and the ground truth labels. It processes the network's predictions and provides corrected supervision signals during training, enabling confusion reduction without permanently modifying the original network architecture or adding complexity to the inference process.
2Measurement precision
If additional modules are added to the neural network to reduce class confusions, then classification accuracy is improved, but runtime increases
Solution Approach 1:
The training head is segmented as a temporary component that exists only during the training phase. It is attached to the neural network for training purposes and can be removed afterward, ensuring that it does not contribute to runtime during inference. This segmentation allows accuracy improvement during training without runtime penalty during deployment.
Solution Approach 2:
The confusion reduction mechanism is applied as a preliminary action during the training phase. The training head processes predictions and provides corrected supervision signals before the final model is deployed. This preliminary intervention during training prepares the network for better classification without adding overhead during actual runtime operations.
3Measurement precision
If loss functions are modified to reduce class confusions, then classification accuracy is improved, but detection performance deteriorates
Solution Approach 1:
The patent separates the classification improvement function from the detection function by introducing a dedicated training head. The original loss function continues to optimize detection performance while the training head introduces an additional supervision signal focused on reducing class confusions. This segmentation allows both objectives to be pursued simultaneously without one compromising the other.
Solution Approach 2:
The patent merges two loss functions: the original detection loss and the additional confusion-reduction loss from the training head. Both loss functions are combined to provide comprehensive supervision that simultaneously optimizes detection accuracy and classification precision. This merging allows the network to learn both detection and classification discrimination without trade-offs.
Data Source
AI summary
A method for training an object detector, an object detector, a computer program and a computer-readable medium. The method includes obtaining a set of object classes confused by the neural network, an object feature map from the neural network, and an object class label comprising object classes assigned to each object in the object feature map. Subsequently, a training head to determine a truthfulness of an object class assigned to an object in an object feature map is added to the neural network. A object class label for which at least one object class assignment is modified is determined. This modified object class label with the object feature map is processed generating an output and compared with a truthfulness label comprising the indication of truthfulness of each object class of the modified object class label by using a first objective function. Network parameters are then updated based on the comparison.