Object Embedding System Jointly Optimizes Detection and Embedding

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image processing systems for object identification often require multiple independent processes for object detection, thumbnailing, and embedding, which can lead to suboptimal performance and increased computational resources, as they do not jointly optimize these processes.

Innovation Solution

A computer-implemented method that trains an object embedding system using a thumbnailing neural network and an embedding neural network, jointly optimizing object detection and embedding processes through iterative training with a loss function that reduces distances between same-instance embeddings and increases distances between different-instance embeddings, allowing for the generation of superior object embeddings with reduced computational resource consumption.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If multiple independent processes are used for object detection, thumbnailing, and embedding, then each process can be independently optimized, but the overall system complexity increases and computational resources are wasted

Engineering Contradiction:
Improveobject identification accuracyVSAvoidsystem complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent combines object detection, thumbnailing, and embedding processes into a single integrated neural network framework. The detection network and thumbnailing network share common components and are trained jointly, eliminating the need for separate independent processes while maintaining or improving overall system performance.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The integrated neural network performs multiple functions simultaneously - it conducts object detection, generates thumbnail representations, and creates embeddings for comparison. This multi-functional approach replaces multiple specialized processes with a single universal system that handles all tasks.

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

2Ease of operation

If multiple independent processes are used for object detection, thumbnailing, and embedding, then each process can operate independently, but computational resource consumption increases

Engineering Contradiction:
Improveindependent process operationVSAvoidcomputational resource consumption
Core Design Contradiction:
Ease of operationVSUse of energy by moving object

Solution Approach 1:

By merging the detection and thumbnailing processes into a single trained framework, the system eliminates redundant computations that would occur in separate processes. The shared neural network components process information once rather than multiple times, reducing overall computational resource consumption.

Inventive Principle:
Principle #5Merging (Combining)

3Ease of manufacture

If traditional separate processes are used for detection and embedding, then implementation is simpler, but embedding quality is suboptimal

Engineering Contradiction:
Improveimplementation simplicityVSAvoidembedding quality
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The system uses a loss function that provides feedback during training to optimize embedding quality. The loss function compares generated embeddings with reference embeddings and adjusts the neural network parameters to minimize the distance between same-instance embeddings while maximizing the distance between different-instance embeddings, thereby improving embedding quality through iterative optimization.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3679525B1Generating object embeddings from images
Publication Date: 2025.03.26 GOOGLE LLC
  • EP3679525B1 patent drawingFigure 1
  • EP3679525B1 patent drawingFigure 2
  • EP3679525B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training an object embedding system. In one aspect, a method comprises providing selected images as input to the object embedding system and generating corresponding embeddings, wherein the object embedding system comprises a thumbnailing neural network and an embedding neural network. The method further comprises backpropagating gradients based on a loss function to reduce the distance between embeddings for same instances of objects, and to increase the distance between embeddings for different instances of objects.