Adaptive Bounding Box Distribution for Fisheye Lens Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Object detection accuracy significantly drops when using a normal lens-trained apparatus on fisheye images due to severe distortion, making it impractical to collect and retrain with one million fisheye images for lens changes.
Innovation Solution
An object detection apparatus and method that determine bounding box distribution based on lens configuration, using a boundary box decision circuit to assign bounding boxes to different detection distances, allowing for accurate object detection on fisheye images without the need for extensive retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a normal lens-trained object detection apparatus is used for fisheye images, then the device complexity remains low, but the object detection accuracy drops significantly
Solution Approach 1:
The patent changes the parameters of bounding box distribution based on lens configuration. Different lens types (normal, fisheye, wide-angle) have different distortion characteristics, and the system adjusts bounding box sizes and positions accordingly to maintain detection accuracy without requiring complete retraining
Solution Approach 2:
The object detection apparatus is designed to handle multiple lens types using a universal training approach. By incorporating lens configuration information and adapting bounding box distributions, the same trained model can work across different lens types, reducing the need for separate training datasets for each lens
2Measurement precision
If one million fisheye images are collected and used for retraining, then the object detection accuracy improves, but the loss of time and resources increases significantly
Solution Approach 1:
The system performs preliminary actions by pre-calculating and storing optimal bounding box distributions for different lens configurations. When a new image is processed, the system quickly retrieves the appropriate bounding box distribution based on the lens type rather than performing extensive retraining, saving significant time and computational resources
Solution Approach 2:
Instead of creating entirely new training data for each lens type, the system copies and adapts the existing bounding box distributions from the universal training model. The bounding box parameters are adjusted based on lens-specific distortion characteristics, avoiding the need to collect and label millions of new images
3Adaptability or versatility
If the bounding box distribution is fixed for normal images, then the manufacturing precision is high, but the adaptability to different lens configurations decreases
Solution Approach 1:
The bounding box distribution is made dynamic rather than fixed. The system automatically adjusts bounding box parameters based on the detected lens configuration, transitioning between different distribution patterns to match the optical characteristics of normal, fisheye, wide-angle, and other lens types while maintaining detection precision
Data Source
AI summary
An object detection apparatus includes a boundary box decision circuit and a processing circuit. The boundary box decision circuit receives lens configuration information of a lens, and refers to the lens configuration information to determine a bounding box distribution of bounding boxes that are assigned to different detection distances with respect to the lens for detection of a target object. The processing circuit receives a captured image that is derived from an output of an image capture device using the lens, and performs object detection upon the captured image according to the bounding box distribution of the bounding boxes.