Object Detection Memory Management via Feature Merging

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing object detection systems in computer vision suffer from low memory efficiency due to overlapping information in key and value pairs and unnecessary storage of similar frames, making them unsuitable for processing high-resolution images.

Innovation Solution

An object detection and classification system that includes an input value generation unit, memory value generation unit, memory management unit, aggregated value generation unit, and object detection unit, which efficiently manages memory by selecting important information, storing only high-importance values, and calculating weighted sums to reduce memory usage and improve accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If key and value are separately used in memory network, then object detection accuracy is maintained, but memory efficiency deteriorates due to information overlap

Engineering Contradiction:
Improveobject detection accuracyVSAvoidmemory usage
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent merges the key and value features into a single unified feature map, eliminating the separate key and value storage structures. This consolidation removes redundant information overlap while preserving the essential features needed for object detection, thereby improving memory efficiency without sacrificing detection accuracy.

Inventive Principle:
Principle #5Merging (Combining)

2Measurement precision

If all past frames are stored in memory, then motion prediction accuracy is improved, but memory usage increases due to storage of similar frames

Engineering Contradiction:
Improvemotion prediction accuracyVSAvoidmemory usage
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent extracts and stores only the essential motion information from past frames rather than storing complete frame data. By taking out only the necessary motion features and temporal differences, the system maintains accurate motion prediction while significantly reducing memory consumption by excluding redundant frame information.

Inventive Principle:
Principle #2Taking out (Extraction)

3Measurement precision

If high resolution images are processed, then detection accuracy is improved, but memory requirements exceed available capacity

Engineering Contradiction:
Improvedetection accuracyVSAvoidmemory capacity
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent segments high-resolution images into smaller patches or regions that can be processed independently with reduced memory requirements. By dividing the large image into manageable segments, the system can process high-resolution content while keeping memory usage within available capacity, maintaining detection accuracy through systematic processing of all segments.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentUS11676372B2Object/region detection and classification system with improved computer memory efficiency
Publication Date: 2023.06.13 KOREA INST OF SCI & TECH
  • US11676372B2 patent drawing
  • US11676372B2 patent drawing
  • US11676372B2 patent drawing

AI summary

The present disclosure relates to an object detection and classification system with higher accuracy and resolution in a less computer memory environment. The system comprises an input value generation unit to receive an input image and generate an input value including feature information; a memory value generation unit to receive a reference image and generate a memory value including feature information; a memory management unit to select information having high importance from the memory values and store in a computer memory; an aggregated value generation unit to compute similarity between the input value and the memory value, calculate a weighted sum to generate an integrated value, and aggregate the integrated value and the input value; and an object detection unit to detect or classify the object from the input image using the aggregated value.