Hybrid TCN and GNN Architecture for Long Video Action Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing machine learning models struggle with performing accurate action segmentation and prediction for long duration, structured videos, especially in automation domains where action classes may share strong visual similarities and have hard-to-differentiate characteristics.

Innovation Solution

A hybrid architecture combining a temporal convolutional network (TCN) and a bi-directional graph neural network (GNN) is employed. The TCN captures long duration dependencies using 1-D convolution operators and dilation rates, while the GNN processes a graph representation of frame-wise features to refine fine-grain per-frame action predictions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a machine learning model is designed to accurately differentiate between action classes with strong visual similarities, then measurement precision improves, but device complexity increases

Engineering Contradiction:
Improveaction prediction accuracyVSAvoidmodel architecture complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The model architecture is segmented into two specialized components: a temporal convolutional network (TCN) for capturing long-duration temporal dependencies and a bi-directional graph neural network (GNN) for modeling fine-grain spatial relationships between video frames. This segmentation allows each component to specialize in specific aspects of action recognition, improving overall precision without requiring a monolithic complex model

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transforms the video data into a graph representation where video frames are nodes and temporal relationships are edges. This dimensional transformation from traditional spatial-temporal tensors to graph structures enables the GNN to capture fine-grain spatial dependencies that are difficult to model in conventional video processing architectures

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If the model processes long duration videos with fine-grain per-frame predictions, then measurement precision improves, but loss of time increases

Engineering Contradiction:
Improvefine-grain action prediction accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The processing pipeline is segmented into two stages: first, the TCN processes temporal dependencies across the entire video duration to generate frame-wise features; second, the GNN processes these features in a graph representation to refine per-frame predictions. This segmentation enables efficient parallel processing while maintaining fine-grain temporal resolution

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The TCN performs preliminary processing by extracting temporal features and generating frame-wise representations before the GNN refines the predictions. This preliminary action reduces the computational burden on the GNN, allowing it to focus on capturing fine-grain spatial relationships without reprocessing the entire video from scratch

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4517585A1Long duration structured video action segmentation
Publication Date: 2025.03.05 INTEL CORP
  • EP4517585A1 patent drawingFigure 1
  • EP4517585A1 patent drawingFigure 2
  • EP4517585A1 patent drawingFigure 3

AI summary

Machine learning models can process a video and generate outputs such as action segmentation assigning portions of the video to a particular action, or action classification assigning an action class for each frame of the video. Some machine learning models can accurately make predictions for short videos but may not be particularly suited for performing action segmentation for long duration, structured videos. An effective machine learning model may include a hybrid architecture involving a temporal convolutional network and a bidirectional graph neural network. The machine learning model can process long duration structured videos by using a temporal convolutional network as a first pass action segmentation model to generate rich, frame-wise features. The frame-wise features can be converted into a graph having forward edges and backward edges. A graph neural network can process the graph to refine a final fine-grain per-frame action prediction.