Bidirectional RNN for Overlapping Video Event Description

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing video description generation methods struggle to accurately distinguish and describe overlapping events due to the use of unidirectional recurrent networks, leading to incomplete and non-distinctive descriptions.

Innovation Solution

The method employs a combination of forward and backward recurrent neural networks to extract and encode video features, enabling accurate positioning and differentiation of overlapping events by generating a forward hidden state sequence and a backward hidden state sequence, which are used to predict event intervals and generate descriptive words.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If a unidirectional recurrent network is used to generate video descriptions, then the model structure is simple and easy to implement, but overlapping events cannot be accurately positioned and distinguished

Engineering Contradiction:
Improvemodel structureVSAvoidevent positioning accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent applies bidirectional recurrent networks that process video frames in both forward and backward directions. The forward recurrent network captures temporal dependencies from past frames, while the backward recurrent network captures dependencies from future frames. This dual-directional approach enables accurate positioning of overlapping events by considering both historical and future contextual information, resolving the limitation of unidirectional networks that can only process sequences in one direction.

Inventive Principle:
Principle #13The other way round (Inversion)

Solution Approach 2:

The patent introduces a temporal dimension by processing video frames at multiple time points (both past and future relative to the current frame). The bidirectional recurrent networks operate in two temporal dimensions simultaneously, allowing the system to distinguish overlapping events by examining their temporal relationships from both directions. This multi-dimensional temporal processing enables accurate event positioning that cannot be achieved with single-directional temporal processing alone.

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

2Productivity

If a unidirectional recurrent network is used for video description generation, then the computational process is straightforward, but events with close end times cannot be accurately positioned

Engineering Contradiction:
Improvecomputational efficiencyVSAvoidevent time positioning accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The bidirectional recurrent network processes video sequences in reverse order as well as forward order. The backward recurrent network specifically processes frames from the end of the video backward to the current frame, allowing the system to accurately determine event end times by examining future frames. This reverse processing capability enables precise temporal boundary detection for events with close end times, overcoming the limitation of unidirectional processing that cannot access future temporal information.

Inventive Principle:
Principle #13The other way round (Inversion)

3Measurement precision

If a bidirectional recurrent neural network is used to encode video features, then overlapping events can be accurately positioned and distinguished, but the model complexity increases

Engineering Contradiction:
Improveevent positioning accuracyVSAvoidmodel structure
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent merges the forward and backward recurrent networks into a unified bidirectional processing framework. Both networks share the same architectural components (recurrent units, weight matrices, bias terms) and process the same video input sequences, just in opposite temporal directions. This merging approach allows the system to achieve accurate event positioning through bidirectional processing while minimizing the increase in model complexity by reusing the same computational building blocks for both forward and backward processing.

Inventive Principle:
Principle #5Merging (Combining)

4Manufacturing precision

If a bidirectional recurrent neural network is used to generate video descriptions, then event descriptions become more distinctive and accurate, but the computational cost increases

Engineering Contradiction:
Improvedescription accuracyVSAvoidcomputational cost
Core Design Contradiction:
Manufacturing precisionVSUse of energy by moving object

Solution Approach 1:

The patent applies bidirectional recurrent networks selectively to process only the portions of the video sequence necessary for event detection and description generation. Rather than processing the entire video sequence bidirectionally at full resolution, the system can focus computational resources on specific time windows or frames where events are detected. This partial application approach maintains high description accuracy for overlapping events while reducing overall computational cost by avoiding unnecessary processing of all video content.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3745305B1Video description generation method and device, video playing method and device, and storage medium
Publication Date: 2023.04.26 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3745305B1 patent drawingFigure 1~2
  • EP3745305B1 patent drawingFigure 3
  • EP3745305B1 patent drawingFigure 4~6

AI summary

Disclosed in the present invention are a video description generation method and device, a video playing method and device, and a computer readable storage medium. The method comprises the following steps: extracting a video feature to obtain a video feature sequence corresponding to a video encoding moment in a video stream; encoding the video feature sequence through a forward recurrent neural network and a backward recurrent neural network to obtain a forward implicit state sequence and a backward implicit state sequence corresponding to each video encoding moment; and according to the forward implicit state sequence and the backward implicit state sequence, locating an event corresponding to each video encoding moment and an interval corresponding to the event at the video encoding moment, thereby predicting a video content description of the event. According to the method of the present invention, overlapping events are distinguished, and the interval corresponding to the event is introduced to predict a word corresponding to the generated event at the video encoding moment. For events that overlapped at the video encoding moment, even if overlap occurs, the overlap still corresponds to different intervals. The video content description of the event at the video encoding moment will have a very high degree of discrimination, and accordingly, there will be a higher degree of discrimination for a given video stream.