Object Detection Model Adjustment for Unknown 3D Point Cloud Objects
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, such as pedestrians or new types of vehicles, due to their reliance on closed training data sets.
Innovation Solution
A method and apparatus for adjusting object detection models by determining object boxes with unknown categories using a combination of three-dimensional point cloud data and bird's-eye view features, incorporating a pseudo-label selection model to filter and train the model with object boxes of unknown categories, thereby improving detection accuracy and versatility.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a closed training data set is used for object detection model training, then the model training process is simple and controlled, but the model cannot accurately identify and categorize unknown objects in dynamic real-world scenarios
Solution Approach 1:
The patent transforms the static closed training set into a dynamic open training set that continuously evolves by incorporating unknown objects detected in real-world scenarios. The model adapts dynamically by integrating newly detected unknown objects with confidence scores into the training data, enabling continuous improvement of detection accuracy for previously unrecognized object categories.
Solution Approach 2:
The patent applies preliminary filtering and confidence scoring to potential unknown objects before incorporating them into the training set. By pre-evaluating detection confidence scores and filtering criteria, the system ensures that only high-quality unknown object candidates are added to the training data, maintaining reliability while expanding adaptability.
2Productivity
If the model is trained only on known object categories, then the training process is efficient and straightforward, but the model lacks versatility to handle diverse and unpredictable real-world scenarios
Solution Approach 1:
The patent implements partial incorporation of unknown objects into the training set by selecting only those with high confidence scores and meeting specific filtering criteria. This partial action approach maintains training efficiency by not incorporating all possible unknown objects, while still progressively expanding the model's versatility to handle diverse real-world scenarios.
Solution Approach 2:
The patent establishes a continuous training loop where the model is repeatedly trained on an expanding open training set that accumulates unknown objects over time. This continuous useful action ensures that the model progressively improves its versatility to detect unknown categories while maintaining efficient training through automated workflows and confidence-based filtering.
3Measurement precision
If the object detection model is adjusted frequently to improve unknown object detection, then the detection accuracy for unknown objects improves, but the system complexity and computational resources increase
Solution Approach 1:
The patent implements a feedback mechanism where detection results including confidence scores are continuously fed back into the training process. Unknown objects detected with high confidence are automatically incorporated into the training set, triggering model retraining. This closed-loop feedback system improves detection accuracy for unknown objects while automating the process to manage system complexity.
Solution Approach 2:
The system performs self-service by automatically detecting unknown objects, evaluating their confidence scores, filtering candidates, and incorporating them into the training set without manual intervention. This self-service approach improves detection accuracy while reducing the operational complexity and resource burden on external operators.
Data Source
AI summary
Methods and related apparatuses for adjusting an object detection model are disclosed. The method includes (i) determining a first object detection result based on three-dimensional point cloud data using a first object detection model, the first object detection result including location information of a first object candidate box, (ii) 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, (iii) determining a second object detection result including 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 (iv) 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.