BEV Object Detection Training with Uncertainty-Based Confidence
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current object detection models face challenges in accurately predicting objects from birds-eye-view (BEV) images generated by sensors like LiDAR and radar, which provide 3D data but only 'see' up to the first point of reflection, leading to information-free areas and reduced prediction accuracy, especially when using centerness approaches.
Innovation Solution
A computer-implemented method for training BEV object detection models that assigns target confidence values to pixels based on uncertainty distributions related to distance, position, rotation, size, and class, allowing the model to learn the importance of pixels within bounding boxes and adjust its parameters to improve detection accuracy and reduce computational overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If sensors like LiDAR or radar are used to provide 3D view and distance information, then depth information and spatial awareness are improved, but the area behind the first reflection point remains unobserved creating information-free areas that reduce detection accuracy
Solution Approach 1:
The patent applies local quality by assigning different confidence values to different pixel regions in the BEV image. Pixels corresponding to areas with reliable sensor data (front reflection points) receive higher confidence values, while pixels in information-free areas (behind reflection points) receive lower confidence values. This allows the model to adaptively weigh different regions based on their information reliability, resolving the contradiction between depth accuracy and information completeness.
Solution Approach 2:
The patent changes the parameter of pixel confidence values from uniform to variable based on spatial position and data reliability. By introducing confidence values as a new parameter that varies across the image, the model can handle the information-free areas without compromising the accuracy of detected objects in well-observed regions.
2Ease of operation
If centerness approach is used for object detection in BEV images, then detection simplicity is improved, but prediction accuracy deteriorates because object centers often fall in information-free areas
Solution Approach 1:
The patent modifies the centerness approach by introducing local confidence values that vary across the image. Instead of treating all pixels equally, the model now considers the reliability of data in different regions, assigning higher weights to pixels with reliable sensor information and lower weights to pixels in information-free areas, thus maintaining simplicity while improving accuracy.
Solution Approach 2:
The patent implements feedback by using the confidence values to guide the detection process. The confidence values provide feedback about the reliability of different regions, allowing the model to adjust its predictions accordingly and compensate for the limitations of the centerness approach in information-free areas.
3Adaptability or versatility
If comprehensive sensor fusion is used to achieve comprehensive detection solution, then detection completeness is improved, but processing complexity increases
Solution Approach 1:
The patent extracts the confidence value assignment as a separate, manageable component from the overall sensor fusion processing. By isolating this function, the model can handle comprehensive sensor data without proportionally increasing processing complexity, as the confidence values are computed based on established spatial relationships and data reliability principles.
Data Source
AI summary
A computer-implemented method for training a birds-eye-view (BEV) object detection model includes inputting a training sample into the model. The training sample includes a BEV image with multiple pixels, and multiple target confidence values. Each pixel of the pixels is associated with a target confidence value of the target confidence values. The method includes receiving as output from the model multiple predicted confidence values. Each predicted confidence value is associated with a pixel of the pixels. The method includes adjusting a parameter set of the model according to a loss. The loss is based on the predicted confidence values and the target confidence values.