Object Detection Model Adjustment for Unknown 3D Object Categories
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing three-dimensional object detection models struggle to accurately identify and categorize unknown objects in dynamic and unpredictable real-world scenarios, limiting their effectiveness in autonomous driving systems.
Innovation Solution
A method and apparatus for adjusting object detection models by combining a first object detection model with a second object detection model using bird's-eye view features, filtering object boxes of unknown categories, and adjusting the second model based on these boxes to improve detection accuracy and versatility.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single object detection model is used, then the device complexity is low, but the detection accuracy for unknown objects deteriorates
Solution Approach 1:
The patent divides the object detection task into two specialized models: a first object detection model for detecting objects in point cloud data, and a second object detection model for detecting objects in bird's-eye view images. This segmentation allows each model to specialize in specific data types and object categories, improving overall detection accuracy while managing complexity through functional division.
Solution Approach 2:
The patent transitions from three-dimensional point cloud data to two-dimensional bird's-eye view images by performing coordinate transformations and generating top-down views of the point cloud data. This dimensional change enables the use of different detection approaches and improves the detection of unknown objects by leveraging complementary information from multiple representations.
2Adaptability or versatility
If the object detection model is trained only on known categories, then the training data requirement is manageable, but the adaptability to unknown objects deteriorates
Solution Approach 1:
The patent performs preliminary filtering to identify candidate object boxes that may contain unknown objects before proceeding to detection. By pre-identifying regions of interest and filtering out obvious known objects, the system prepares the data in advance for specialized processing, enabling better adaptability to unknown objects without requiring excessive training data.
Solution Approach 2:
The patent introduces an intermediary filtering mechanism that identifies candidate boxes for unknown objects based on the outputs of both detection models. This intermediary step acts as a bridge between the two models, selecting regions that require further investigation and enabling the system to adapt to unknown objects without needing comprehensive training data for all possible categories.
3Measurement precision
If multiple detection models are combined, then the detection accuracy improves, but the computing resource consumption increases
Solution Approach 1:
The patent applies partial action by using the two detection models in a coordinated manner rather than running them independently for all objects. The first model processes point cloud data while the second model processes bird's-eye view images, and their results are combined only where needed, reducing redundant computations while maintaining high detection accuracy.
Solution Approach 2:
The patent merges the detection results from both models by combining their outputs to identify unknown objects. Rather than running both models separately and independently, the system integrates their findings, using the strengths of each model to compensate for the other's limitations, thereby improving accuracy while optimizing resource utilization through coordinated processing.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present disclosure related to methods and related apparatuses for adjusting an object detection model. The method comprises determining a first object detection result based on three-dimensional point cloud data using a first object detection model, the first object detection result comprising location information of a first object candidate box. The method further comprises determining, from the first object candidate box, a first object box, the category of objects contained in the first object box being a category unknown to the second object detection model. The method further comprises determining a second object detection result comprising location information of a second object box and a category of objects contained in the second object box using a second object detection model based on the bird's-eye view features associated with the three-dimensional point cloud data and adjusting the second object detection model based on the second object box and the first object box. In this way, the object detection model can be adjusted using the object boxes determined by other object detection models, thereby improving the model's detection performance for objects of unknown categories.