Point Cloud Object Detection With Annotation-Efficient Sample Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection methods in point cloud data face challenges with high annotation costs and limited dataset effectiveness due to complex data enhancement processes and random placement of annotation frames, leading to reduced detection accuracy.
Innovation Solution
A method involving the extraction of instance point cloud data using annotation information, determining an object position point, and fusing raw and instance point cloud data to create a fused to-be-detected sample, which is then used for object detection, thereby increasing sample size and improving detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If random placement of annotation frames is used to generate enhancement data, then data generation simplicity is improved, but detection accuracy deteriorates due to lack of validity in generated training dataset
Solution Approach 1:
The patent applies preliminary action by pre-processing the point cloud data to identify valid placement positions for annotation frames before actually placing them. This involves analyzing the scene geometry and determining positions where objects can be realistically added, ensuring that the data enhancement maintains physical plausibility and improves detection accuracy while keeping the process automated and simple.
2Measurement precision
If structured data enhancement is used to generate training data set, then detection accuracy is improved, but process complexity increases and online data enhancement becomes difficult to achieve
Solution Approach 1:
The patent implements self-service by enabling the system to automatically generate its own training data without requiring external structured data sources or complex manual processing. The method uses the existing point cloud data and automatically determines valid placement positions for annotation frames, allowing the system to enhance its own training dataset online without human intervention or complex external processes.
3Measurement precision
If annotation data size is increased to improve model performance, then detection accuracy is improved, but annotation cost and time increase
Solution Approach 1:
The patent applies copying by creating synthetic copies of objects through annotation frames and placing them at valid positions in the point cloud data. Instead of manually annotating each training sample, the system copies existing object models and instances, automatically generates their point cloud representations, and places them in valid positions, thereby dramatically increasing the training dataset size without proportional increases in annotation time or cost.
Data Source
AI summary
An object detection method is provided. In the method, raw point cloud data including a to-be-detected object is obtained, where the raw point cloud data includes annotation information for the to-be-detected object; instance point cloud data corresponding to the to-be-detected object is extracted from the raw point cloud data by using the annotation information; an object position point is determined from the raw point cloud data, and the raw point cloud data and the instance point cloud data to obtain a fused to-be-detected sample is fused based on the object position point and the to-be-detected object is detected by using the raw point cloud data and the to-be-detected sample.