Ad-hoc Event Detection via Pre-computed Label Distributions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic detection and recognition systems for objects and events in multimedia content require large amounts of training data and resources, making them costly and inefficient.
Innovation Solution
A system creates an event label distribution using a combination of textually and visually based label distributions to detect objects associated with semantic events, allowing for ad-hoc detection by comparing relevance scores to determine the type of event present in multimedia content.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large quantities of training data are used to train automatic detection and recognition systems, then detection accuracy is improved, but resource consumption and training time increase significantly
Solution Approach 1:
The system performs preliminary action by pre-computing and storing label distributions for training data before actual event detection is needed. This allows the detection phase to use these pre-computed distributions directly, avoiding the need to re-process large training datasets during detection, thus reducing resource consumption while maintaining detection accuracy
Solution Approach 2:
The training data and label distributions are segmented into event-specific distributions that can be independently stored and reused. Instead of processing the entire training dataset for each detection task, the system segments and utilizes only the relevant event label distributions, reducing computational resources required during detection
2Reliability
If traditional training-based detection systems are used, then event detection capability is achieved, but training time and computational resources are prohibitively expensive
Solution Approach 1:
Label distributions for all training events are computed and stored in advance before actual detection tasks. This preliminary computation eliminates the need for time-consuming training processes during deployment, allowing the system to perform rapid event detection by comparing query events against pre-computed distributions
Solution Approach 2:
Instead of re-training or re-processing original training data during detection, the system creates and uses copies in the form of pre-computed label distributions. These distributed representations capture the essential characteristics of training events without requiring access to the original training datasets, enabling fast detection with minimal computational overhead
Data Source
AI summary
Event detection may be carried out to determine one or more content items associated with an event type. A server may determine one or more visually relevant labels, of a plurality of labels, associated with the event type. The server may determine one or more textually relevant labels, of the plurality of labels, associated with the event type. The server may determine an event label distribution based on a correlation of the visually relevant event labels and the textually relevant labels, wherein the event label distribution comprises a plurality of labels. The server may determine a content item label distribution associated with a content item, wherein the content item label distribution comprises a plurality of second labels. The server may send, based on comparing the event label distribution with the content item label distribution, the content item to a computing device.