Memory-Guided Video Object Detection for Mobile Real-Time Accuracy

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing video object detection models struggle to achieve real-time performance on mobile devices without significant accuracy loss, as they are constrained by computation and energy limitations.

Innovation Solution

An interleaved object detection model utilizing multiple feature extractor networks and a shared memory layer, where one network prioritizes accuracy and the other prioritizes speed, with an adaptive or fixed interleaving policy to optimize frame analysis, allowing for efficient and accurate object detection.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a single elaborate convolutional neural network design is used to improve accuracy, then object detection accuracy improves, but processing speed and real-time performance deteriorate

Engineering Contradiction:
Improveobject detection accuracyVSAvoidprocessing speed
Core Design Contradiction:
Measurement precisionVSSpeed

Solution Approach 1:

The patent divides the object detection task into multiple sub-tasks by using separate feature extractor networks (e.g., ResNet for accuracy, MobileNet for speed) that process different aspects of the input image. Each network is specialized for specific detection requirements, allowing the system to achieve both high accuracy and real-time performance by segmenting the computational workload rather than using a single monolithic network.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements dynamic selection between different feature extractor networks based on the specific detection task requirements. The system can switch between accurate but slow networks and fast but less accurate networks depending on the temporal and computational constraints, enabling adaptive performance optimization that resolves the accuracy-speed tradeoff dynamically rather than statically.

Inventive Principle:
Principle #15Dynamics

2Productivity

If neural networks are deployed on mobile devices with extreme computation and energy constraints, then portability and real-time operation improve, but detection accuracy deteriorates

Engineering Contradiction:
Improvereal-time operation capabilityVSAvoiddetection accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent applies partial action by using lightweight feature extractor networks (such as MobileNet or EfficientNet with reduced complexity) that perform sufficient detection at lower computational cost. Instead of always using the most accurate but computationally expensive network, the system uses partial models that provide adequate accuracy for mobile device constraints while enabling real-time operation, accepting that maximum accuracy is not always achieved but sufficient performance is maintained.

Inventive Principle:
Principle #16Partial or excessive action

3Speed

If multiple feature extractor networks are used with different accuracy-speed tradeoffs, then real-time performance improves, but model complexity and computational cost increase

Engineering Contradiction:
Improvereal-time performanceVSAvoidmodel complexity
Core Design Contradiction:
SpeedVSDevice complexity

Solution Approach 1:

The patent implements multi-functionality by designing a unified detection framework that can accommodate multiple feature extractor networks with different characteristics. The system uses a common detection head and shared computational graph that can work with various backbone networks (ResNet, MobileNet, EfficientNet), allowing one detection system to serve multiple accuracy-speed requirements without requiring separate complete detection pipelines for each network type, thus managing complexity through universality.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP3928247B1Memory-guided video object detection
Publication Date: 2026.04.08 GOOGLE LLC
  • EP3928247B1 patent drawingFigure 1
  • EP3928247B1 patent drawingFigure 2
  • EP3928247B1 patent drawingFigure 3

AI summary

Systems and methods for detecting objects in a video are provided. A method can include inputting a video comprising a plurality of frames into an interleaved object detection model comprising a plurality of feature extractor networks and a shared memory layer. For each of one or more frames, the operations can include selecting one of the plurality of feature extractor networks to analyze the one or more frames, analyzing the one or more frames by the selected feature extractor network to determine one or more features of the one or more frames, determining an updated set of features based at least in part on the one or more features and one or more previously extracted features extracted from a previous frame stored in the shared memory layer, and detecting an object in the one or more frames based at least in part on the updated set of features.