Rotation-Invariant Depth Map Features for Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computer vision systems face challenges in accurately identifying objects like body parts due to their varying orientations, as existing features are not invariant to rotations, leading to poor accuracy and increased complexity in classifiers.
Innovation Solution
The development and use of rotation-invariant features that can be applied to depth maps, allowing for the estimation of local orientations and transformation of feature regions, reducing the need for extensive training data and simplifying the classifier without compromising accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If example images contain many different rotations to account for all possible orientations, then the accuracy of the classifier is improved, but the complexity of the classifier increases leading to slow processing speed and high memory usage
Solution Approach 1:
The patent applies parameter changes by transforming features into a rotation-invariant representation. Specifically, it estimates local orientation for each depth pixel and transforms feature regions from the image coordinate system to a local coordinate system that rotates with the object. This allows the classifier to recognize objects regardless of their rotation angle, improving accuracy without requiring multiple rotated training examples, thus reducing classifier complexity
Solution Approach 2:
The patent introduces an additional dimensional transformation by adding orientation estimation and coordinate system transformation steps. Instead of handling rotation through more training data (increasing complexity), it transforms the feature extraction process into a new dimensional space (local coordinate system) where rotation invariance is naturally achieved, resolving the contradiction between accuracy and complexity
2Ease of manufacture
If features are not invariant to object orientations, then fewer example images are needed for training, but the accuracy of object identification deteriorates when objects appear in different orientations
Solution Approach 1:
The patent changes the parameter of feature representation from fixed image-coordinate-based features to rotation-invariant local-coordinate-based features. By estimating local orientation and transforming features accordingly, the system achieves orientation invariance, allowing accurate identification of objects in any orientation while maintaining ease of training with fewer examples
3Productivity
If the classifier is simplified to reduce processing requirements, then processing speed is improved, but the ability to handle multiple object orientations is reduced
Solution Approach 1:
The patent makes the feature extraction process universal by creating rotation-invariant features. The local coordinate system transformation allows the same simplified classifier to handle objects in any orientation, giving it multi-functionality. A single classifier structure can now process objects regardless of rotation, improving processing speed while maintaining adaptability to various orientations
Data Source
AI summary
Technology is described for determining and using invariant features for computer vision. A local orientation may be determined for each depth pixel in a subset of the depth pixels in a depth map. The local orientation may an in-plane orientation, an out-out-plane orientation or both. A local coordinate system is determined for each of the depth pixels in the subset based on the local orientation of the corresponding depth pixel. A feature region is defined relative to the local coordinate system for each of the depth pixels in the subset. The feature region for each of the depth pixels in the subset is transformed from the local coordinate system to an image coordinate system of the depth map. The transformed feature regions are used to process the depth map.