Object-Centric Pose Estimation With Dynamic Keyframe Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing scanning processes generate an unnecessarily large number of keyframes in pose graphs, leading to inefficiency and resource wastage, particularly in scenarios with limited computing resources, due to selecting keyframes based on six degrees of freedom criteria.
Innovation Solution
A method for selecting keyframes based on two degrees of freedom criteria, ensuring redundancy reduction by choosing frames that maintain the object in view and using pose optimization, with additional frames selected if the entire object is not visible, thereby optimizing keyframe selection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If keyframes are selected based on six degrees of freedom criteria, then comprehensive coverage of object views is achieved, but the number of keyframes becomes unnecessarily large
Solution Approach 1:
The patent extracts and removes redundant keyframes from the pose graph by identifying frames that provide overlapping or redundant views of the object. Specifically, it removes keyframes where the camera pose differs only by rotation around the object center (yaw/pitch/roll) when those rotations do not provide new information about the object's appearance, thereby reducing the total number of keyframes while maintaining comprehensive coverage.
Solution Approach 2:
The patent changes the parameter used for keyframe selection from six degrees of freedom (6DoF) to two degrees of freedom (2DoF) by considering only the camera's position relative to the object center in 3D space, rather than all possible rotational orientations. This parameter change fundamentally reduces the search space for redundant keyframes while maintaining sufficient view coverage.
2Measurement precision
If more keyframes are selected to ensure accurate object representation, then localization accuracy is improved, but computational efficiency deteriorates
Solution Approach 1:
The patent extracts and removes computationally expensive pose optimization operations from frames that are determined to be redundant. By identifying and removing keyframes that do not contribute new information about the object, the system maintains localization accuracy using a reduced set of essential keyframes, thereby improving overall computational efficiency.
Solution Approach 2:
The patent changes the selection criteria from 6DoF to 2DoF, which reduces the number of keyframes that require computationally intensive pose optimization. This parameter change directly reduces the total computational load while maintaining sufficient accuracy for localization tasks.
3Loss of information
If keyframes are selected using traditional six degrees of freedom criteria, then complete pose information is captured, but resource utilization becomes inefficient
Solution Approach 1:
The patent extracts and removes redundant keyframes that do not provide essential pose information for accurate localization. By identifying frames where the camera's 2D position on the object's surface is identical to existing keyframes, the system removes unnecessary computational resources while preserving all critical pose information needed for accurate object representation.
Solution Approach 2:
The patent changes from tracking six degrees of freedom to tracking only two degrees of freedom (camera position on object surface), which fundamentally reduces the amount of pose information that needs to be stored and processed. This parameter change eliminates redundant computational resources while maintaining the essential information needed for localization.
Data Source
AI summary
An exemplary process acquires frames of image data in a physical environment including an object, tracks an orientation of the device with respect to the object during the acquiring of the frames, selects keyframes from the frames based on detecting a first condition and selecting a first frame as a keyframe based on a first keyframe selection criterion and the orientation of the device at the first frame, and based on detecting a second condition, selecting a second frame as a keyframe based on a second keyframe selection criterion and the orientation of the device at the second frame, wherein the second keyframe selection criterion is different than the first keyframe selection criterion, and generates an object scan (e.g., a pose graph) based on the frames of image data and the selected keyframes.