Object Recognition Model Training for Visual Tracking Accuracy

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Traditional image processing models, designed for image classification tasks, are not suitable for visual object tracking, resulting in low accuracy and real-time performance in visual object tracking scenarios.

Innovation Solution

A first object recognition model, such as a VGG or ResNet model, is trained and combined with a tracking algorithm to accurately extract features for visual object tracking, using a method that involves acquiring and processing images with marking information to determine the position and size of tracked objects, and updating the model based on difference information to improve feature extraction and tracking performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a traditional image processing model designed for image classification is used, then the model structure is simple and easy to implement, but the tracking accuracy and real-time performance are poor

Engineering Contradiction:
Improvetracking accuracyVSAvoidmodel complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent changes the training parameters and objective function of the object recognition model from image classification to visual tracking. Specifically, it uses tracking-specific loss functions and training data annotations to retrain the model, transforming it into a tracker that optimizes for tracking accuracy rather than classification performance.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent copies and adapts existing object recognition models (such as VGG or ResNet) by retraining them with tracking-specific objectives. It uses pre-trained classification models as starting points and fine-tunes them with tracking loss functions and tracking annotation data, effectively copying the architectural framework while changing the functional behavior.

Inventive Principle:
Principle #26Copying

2Measurement precision

If a deeper network is used to extract rich visual features, then the feature extraction capability is improved, but the processing speed decreases

Engineering Contradiction:
Improvefeature extraction accuracyVSAvoidprocessing speed
Core Design Contradiction:
Measurement precisionVSSpeed

Solution Approach 1:

The patent applies partial action by using shallower network architectures (such as VGG-16 or ResNet-50) rather than extremely deep networks. It achieves sufficient tracking performance with moderate network depth by optimizing the training objective and using tracking-specific loss functions, avoiding the diminishing returns and computational overhead of overly deep networks.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The patent changes the training parameters and objective functions to prioritize tracking performance over raw feature extraction capability. By using tracking-specific loss functions and training data, the model achieves effective tracking with shallower networks that process images faster while maintaining sufficient feature quality for accurate tracking.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3971772B1Model training method and apparatus, and terminal and storage medium
Publication Date: 2023.08.09 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3971772B1 patent drawingFigure 1a
  • EP3971772B1 patent drawingFigure 1b~2
  • EP3971772B1 patent drawingFigure 3a

AI summary

Disclosed in embodiments of the present application are a model training method and apparatus, and a terminal and a storage medium. The method comprises: obtaining a template image and a test image; calling a first object recognition model to process the features of a tracking object in the template image to obtain a first reference response, and calling a second object recognition model to process the features of the tracking object in the template image to obtain a second reference response; calling the first object recognition model to process the features of the tracking object in the test image to obtain a first test response, and calling the second object recognition model to process the features of the tracking object in the test image to obtain a second test response; tracking the first test response to obtain a tracking response of the tracked object; and updating the first object recognition model based on difference information between the first reference response and the second reference response, difference information between the first test response and the second test response, and difference information between a tracking tag and a tracking response.