Video Partitioning Using Bidirectional RNN for Scene and Chapter Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional video partitioning methods can only partition videos into shots, limiting their application in scenarios requiring scene or chapter-level partitions, and often rely solely on visual features, leading to less accurate partitions.

Innovation Solution

A method that uses a bidirectional recurrent neural network to generate fixed-size feature representations of shots, integrating both visual and audio features to determine video partitions into scenes and chapters, allowing for more accurate and efficient partitioning by processing shots rather than frames.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional video partitioning methods process videos frame-by-frame to detect boundaries, then they can identify shot boundaries, but they cannot accurately partition videos into scenes or chapters and require excessive computational resources

Engineering Contradiction:
Improvepartitioning accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the video processing task by shifting from frame-level analysis to shot-level analysis. It first partitions the video into shots using conventional methods, then processes each shot as a discrete unit to extract features and determine scene/chapter boundaries. This segmentation reduces the overall computational complexity while maintaining or improving partitioning accuracy for higher-level structures.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces shot-level feature representations as an intermediary between raw frames and final scene/chapter partitions. Instead of directly analyzing frames to detect scene boundaries, the system first creates compressed shot representations that capture essential visual and audio information, then uses these intermediaries to make partitioning decisions, reducing computational burden.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Adaptability or versatility

If conventional methods partition videos only into shots, then they maintain simple processing, but they limit application in scenarios requiring scene or chapter-level partitions

Engineering Contradiction:
Improvepartitioning capabilityVSAvoidprocessing complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent extends the partitioning capability from one dimension (shot boundaries) to multiple dimensions (scenes and chapters). It implements a hierarchical partitioning system where videos are first divided into shots, then groups of shots are organized into scenes based on visual continuity, and finally scenes are grouped into chapters based on semantic relationships, enabling versatile partitioning without excessive complexity.

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

Solution Approach 2:

The patent implements dynamic partitioning by allowing the system to adaptively determine the appropriate granularity level based on the video content and application requirements. The bidirectional RNN dynamically processes shot sequences to identify scene and chapter boundaries, enabling the system to flexibly partition videos at different levels (shots, scenes, or chapters) depending on the specific application needs.

Inventive Principle:
Principle #15Dynamics

3Measurement precision

If methods rely solely on visual features for partitioning, then they simplify processing, but they achieve less accurate partitions compared to multi-modal approaches

Engineering Contradiction:
Improvepartitioning accuracyVSAvoidfeature processing complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent merges visual and audio feature processing by extracting features from both modalities simultaneously and feeding them into a unified bidirectional RNN model. The visual features come from CNN processing of video frames, while audio features are extracted from the audio track, and both are combined to make partitioning decisions, improving accuracy through multi-modal integration.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The bidirectional recurrent neural network serves as a universal processor that handles both visual and audio features. This single model performs multiple functions: processing visual feature sequences, processing audio feature sequences, and integrating both modalities to detect scene and chapter boundaries, demonstrating multi-functionality that improves partitioning accuracy without proportionally increasing complexity.

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

Data Source

PatentEP3610421B1Partitioning videos
Publication Date: 2024.08.07 GOOGLE LLC
  • EP3610421B1 patent drawingFigure 1A
  • EP3610421B1 patent drawingFigure 1B
  • EP3610421B1 patent drawingFigure 2

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for partitioning videos. In one aspect, a method includes obtaining a partition of a video into one or more shots. Features are generated for each shot, including visual features and audio features. The generated features for each shot are provided as input to a partitioning neural network that is configured to process the generated features to generate a partitioning neural network output. The partition of the video into one or more chapters is determined based on the partitioning neural network output, where a chapter is a sequence of consecutive shots that are determined to be taken at one or more locations that are semantically related.