Quantized Scale-Invariant Feature Embeddings for Mobile Image Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Augmented reality applications on mobile devices face challenges in efficiently transmitting and processing scale-invariant image features due to limited resources, with existing methods being cumbersome for evolving databases and requiring high bit rates.
Innovation Solution
The method involves extracting scale-invariant features from images, performing random projections, and quantizing them to reduce dimensionality, allowing for efficient transmission and search in a low-dimensional subspace, balancing the number of projections and quantization levels to achieve accurate image retrieval with minimal bandwidth.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If SIFT features are used for image matching, then matching accuracy is improved, but transmission bandwidth requirement increases prohibitively
Solution Approach 1:
The patent extracts only the essential components needed for matching by projecting SIFT features onto a lower-dimensional subspace using random projections. This extracts the core matching information while discarding redundant data, achieving accurate matching with significantly reduced bandwidth transmission.
Solution Approach 2:
The patent transforms high-dimensional SIFT feature vectors into low-dimensional representations through random projections onto a subspace. This dimensionality reduction maintains the essential matching characteristics while dramatically reducing the bandwidth required for transmission from the mobile device to the server.
2Measurement precision
If training-based methods are used for compressing image descriptors, then compression accuracy is improved, but system complexity and retraining requirements increase for evolving databases
Solution Approach 1:
The patent employs random projections that do not require training or adaptation to the database content. The method is self-service in the sense that it works universally without needing to learn from or adapt to the specific characteristics of the database, eliminating retraining requirements when the database evolves.
Solution Approach 2:
The patent changes the approach from learning-based parameter optimization to fixed random projection parameters. By using predetermined random projection matrices instead of trained descriptors, the system achieves compression without the complexity of training mechanisms while maintaining effectiveness.
Data Source
AI summary
Scale-invariant features are extracted from an image. The features are projected to a lower dimensional random projection matrix by multiplying the features by a matrix of random entries. The matrix of random projections is quantized to produce a matrix of quantization indices, which form a query vector for searching a database of images to retrieve metadata related to the image.