Neural Network Rotation Hypotheses for 6D Pose Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting the 6D pose of objects, particularly symmetrical objects and dealing with occlusions, are inadequate due to their inability to handle ambiguities and symmetries effectively.
Innovation Solution
A neural network is configured to generate multiple feature maps with varying resolutions and depths, performing classification to predict object types, positions, shapes, and a plurality of rotation hypotheses, which are then refined through clustering and outlier rejection to estimate the 6D pose, using quaternions to model spatial rotations and incorporating principal component analysis to address ambiguities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If a neural network predicts a single rotation hypothesis or pose, then the computational complexity is reduced and the processing speed is improved, but the ability to handle symmetrical objects and occlusions deteriorates
Solution Approach 1:
The patent segments the rotation prediction into multiple discrete rotation hypotheses rather than a single continuous prediction. The neural network outputs a set of candidate rotation angles (e.g., 0°, 45°, 90°, 135°) for each detected object, allowing the system to handle symmetrical objects by evaluating multiple possible orientations simultaneously. This segmentation enables the system to resolve ambiguities in symmetrical objects and occluded objects without significantly increasing computational complexity.
2Measurement precision
If multiple rotation hypotheses are obtained for each object, then the accuracy of 6D pose estimation for symmetrical objects is improved, but the computational complexity and processing time increase
Solution Approach 1:
The patent applies partial action by predicting a limited set of discrete rotation hypotheses (e.g., 4-8 candidate angles) rather than all possible continuous rotations. This partial enumeration of rotation possibilities provides sufficient accuracy for symmetrical objects while keeping computational complexity manageable. The system predicts rotation hypotheses at specific angular intervals rather than exhaustively searching all possible orientations, achieving a balance between precision and computational efficiency.
3Reliability
If the neural network processes images to detect 6D pose with high accuracy, then the reliability is improved, but the processing speed decreases
Solution Approach 1:
The patent performs preliminary actions by pre-defining a set of discrete rotation hypotheses before processing each image. The neural network is trained to output predictions from a predetermined set of rotation angles, which eliminates the need for real-time computation of all possible rotations. This preliminary preparation of rotation candidates allows the system to maintain high accuracy while processing images at a faster rate, as the network only needs to evaluate a fixed set of hypotheses rather than compute the full rotation space for each image.
Data Source
AI summary
A system and a method for processing an image include inputting the image to a neural network configured to: obtain a plurality of feature maps, each feature map having a respective resolution and a respective depth, perform a classification on each feature map to deliver, for each feature map: the type of at least one object visible on the image, the position and shape in the image of at least one two-dimensional bounding box surrounding the at least one object, a plurality of rotation hypotheses for the at least one object.