Distance-Space Embedding for Multi-Descriptor Video Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Video-based object and activity recognition systems face challenges in comparing and combining multiple descriptor types, as no single descriptor can represent both articulated and non-articulated activities, leading to complex processing steps and inefficient retrieval.
Innovation Solution
The use of distance-space embedding to match and retrieve objects from video data by generating descriptors that exist in different spaces, allowing for similarity-based indexing and refinement through user feedback, and relevance weighting, enabling the comparison and combination of heterogeneous descriptor types.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multiple descriptor types are used to represent different activity categories, then the system can handle both articulated and non-articulated activities, but the complexity of comparing and combining descriptors increases
Solution Approach 1:
The patent introduces a distance space as an intermediary representation that mediates between different descriptor types. Each descriptor is transformed into a distance representation in this unified space, allowing comparison and combination of heterogeneous descriptors without direct complexity. The distance space acts as the mediator that enables multi-descriptor fusion while maintaining simplicity in the comparison process.
Solution Approach 2:
The patent transforms descriptors from their original parameter spaces into a unified distance space representation. By changing the representation parameters of descriptors into distance metrics within a common space, the system enables comparison of previously incomparable descriptors. This parameter transformation resolves the contradiction by providing a unified representation that simplifies comparison while maintaining versatility.
2Adaptability or versatility
If retrieval is performed on individual descriptors and results are combined, then the system can handle heterogeneous descriptors, but the retrieval efficiency decreases
Solution Approach 1:
The patent merges the retrieval process for multiple descriptors into a single unified retrieval operation in the distance space. Instead of performing separate retrieval for each descriptor type and then combining results, the system performs one retrieval operation that simultaneously considers all descriptor types. This merging approach maintains the ability to handle heterogeneous descriptors while significantly improving retrieval efficiency by eliminating multiple sequential operations.
Data Source
AI summary
A process includes receiving an archive of video data comprising a plurality of archive descriptor types, and applying a query to the archive. The query includes a number of N query descriptor types for a query object. The process further includes determining a difference between each query descriptor type and corresponding descriptor types of the archive, and storing each difference as a point in an N dimensional space. The process further includes identifying an archive object that is similar to the query object as a function of proximities of the differences to an origin of the N dimensional space.