Federated Video Analytics Models for Unseen Distribution Shifts

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing video analytics models struggle with inefficiency in new or unseen scenarios due to prioritizing variant features over invariant features, leading to out-of-distribution generalization issues and high computational intensity in data generation, particularly in Generative Adversarial Networks (GANs).

Innovation Solution

A federated learning-based approach with a parent-child model architecture is employed, utilizing adaptive weight transfer and knowledge distillation to train models on invariant features across multiple datasets with different underlying distributions, leveraging pre-trained 3DCNN models for feature extraction and hierarchical clustering to generate a robust video analytics model.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If existing models prioritize variant features for easy learning, then model training is simplified, but model performance deteriorates in new or unseen scenarios with different underlying distributions

Engineering Contradiction:
Improvemodel training simplicityVSAvoidmodel performance in unseen scenarios
Core Design Contradiction:
Ease of manufactureVSAdaptability or versatility

Solution Approach 1:

The model is segmented into multiple specialized models, each trained on a specific subset of data with particular underlying distributions. Instead of one general model trying to learn all variants, the system divides the problem into multiple focused models that can be combined through ensemble methods, achieving both specialization and generalization.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system dynamically adjusts model parameters and architecture based on the characteristics of unseen data distributions. By detecting distribution shifts and adapting model parameters accordingly, the system maintains high performance across diverse scenarios without retraining from scratch.

Inventive Principle:
Principle #35Parameter changes

2Adaptability or versatility

If Generative Adversarial Networks are used for data generation, then training data diversity is improved, but computational intensity increases significantly

Engineering Contradiction:
Improvetraining data diversityVSAvoidcomputational intensity
Core Design Contradiction:
Adaptability or versatilityVSUse of energy by moving object

Solution Approach 1:

Data generation and augmentation operations are performed in advance during the training phase to create diverse training datasets. By pre-generating diverse data samples before model training, the system avoids the need for real-time data generation during inference, significantly reducing computational intensity during deployment while maintaining data diversity benefits.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4675576A1System and method for generating a federated learning based model for improved video analytics
Publication Date: 2026.01.07 MOKSA AI
  • EP4675576A1 patent drawingFigure 1
  • EP4675576A1 patent drawingFigure 2
  • EP4675576A1 patent drawingFigure 3A

AI summary

A system and method for generating federated learning-based model for improved video analytics is provided. A set of datasets are fetched from an input unit that comprises video data associated with domain stored in the input unit. Features are extracted from the pre-defined number of frames with respect to an event captured in the video frames by employing a pre-trained model on pre-processed video frames. Video data is passed through the pre-trained model to obtain video-specific features in the form of vectors. Video data is clustered with similar underlying distributions in the same cluster. Lastly, parent-child model architecture is generated. Parent model data is initialized for parent model generation by aggregating a first pre-defined percentage of video data from each cluster in stratified manner. The child models are initialized after generating parent model by using weights of the parent model for predicting event associated with the domain.