Generative Model for Few-Shot Object Recognition

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Object categorization in digital images is challenging due to variations in shape and color among objects within the same category, particularly for novel classes with limited data representation.

Innovation Solution

A computer vision method using a generative model and discriminative model framework, such as StackGAN, to generate and classify digital images based on textual descriptions, leveraging multi-modal training samples to bridge the gap between text and image spaces, and employing adversarial training to improve image quality and classification accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If a classifier is trained on limited data from novel classes, then training time and computational resources are reduced, but classification accuracy deteriorates due to insufficient data representation

Engineering Contradiction:
Improvetraining efficiencyVSAvoidclassification accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The method performs preliminary actions by pre-training the generative model on large datasets of base classes before attempting to classify novel classes. This pre-training establishes a robust feature representation that can be leveraged for few-shot classification, allowing the system to achieve good accuracy even with limited novel class data.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The generative model creates synthetic copies or pseudo-samples of novel class images by translating text descriptions into generated images. These generated images serve as additional training samples, effectively copying the appearance and characteristics of novel classes without requiring actual labeled images, thereby improving classification accuracy with minimal real data.

Inventive Principle:
Principle #26Copying

2Measurement precision

If more data samples are collected for novel classes to improve classification accuracy, then measurement precision improves, but loss of time and resources for data collection and annotation increases

Engineering Contradiction:
Improveclassification accuracyVSAvoiddata collection and annotation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

Instead of collecting and annotating real images for novel classes, the system copies visual information from text descriptions through the generative model. This creates synthetic training samples that capture the essential characteristics of novel classes without requiring time-consuming data collection and manual annotation processes.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The text description serves as an intermediary between the novel class concept and the visual representation. Rather than directly collecting images, the system uses text as a mediator that the generative model translates into synthetic images, significantly reducing the time and resources needed for data preparation.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Device complexity

If traditional classification methods are used on limited novel class data, then device complexity is reduced, but classification accuracy deteriorates due to inability to handle few-shot scenarios

Engineering Contradiction:
Improvesystem simplicityVSAvoidfew-shot classification accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The generative model serves multiple functions: it acts as both a feature extractor and a data generator. By training on base classes and then generating samples for novel classes, the same model performs both representation learning and few-shot classification support, reducing the need for separate complex systems while improving accuracy.

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

Solution Approach 2:

The system changes the parameter of data availability by generating synthetic samples that increase the effective number of training samples for novel classes. This parameter change allows the classifier to achieve better accuracy in few-shot scenarios without fundamentally changing the classification architecture or increasing system complexity.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3754549B1A computer vision method for recognizing an object category in a digital image
Publication Date: 2021.09.22 SAP SE
  • EP3754549B1 patent drawingFigure 1~2
  • EP3754549B1 patent drawingFigure 3~4
  • EP3754549B1 patent drawingFigure 5

AI summary

The present disclosure relates to a method of generating a training data set for training a classifier comprising: providing a first set of multi-modal training samples comprising first visual data samples and first textual data samples of a number of base classes. A second set of multi-modal training samples comprising second visual data samples and second textual data samples of a number of novel classes may be provided. The number of training samples per base class in the first set of multi-modal training samples being greater than the number of training samples per novel class in the second set of multi-modal training samples. A generative model G may be trained using the first set of multi-modal training samples. The trained generative model G may be used to transform the second textual data samples into output visual data samples of the novel classes. The training data set comprises at least part of the output visual data samples.