CRNN Training With Optical Flow State Warping for Video Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for semantic video segmentation struggle with temporal inconsistency and flickering effects due to the lack of effective integration of temporal information, leading to unsatisfactory performance in real-world applications like autonomous driving and augmented reality.

Innovation Solution

A method for training a Convolutional Recurrent Neural Network (CRNN) that incorporates optical flow-based warping of internal states between frames to adapt to pixel motion, using a recurrent module such as FlowingGRU or FlowingLSTM, enhancing temporal consistency in semantic segmentation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If image-level semantic segmentation is performed independently on each frame, then processing speed is maintained, but temporal consistency deteriorates causing flickering effects

Engineering Contradiction:
Improveprocessing speedVSAvoidtemporal consistency
Core Design Contradiction:
SpeedVSReliability

Solution Approach 1:

The patent applies continuity of useful action by maintaining temporal continuity in semantic segmentation through recurrent neural networks. The RNN processes video frames sequentially, preserving temporal dependencies and ensuring consistent segmentation across frames, thereby eliminating flickering effects while maintaining reasonable processing speed through efficient recurrent computation.

Inventive Principle:
Principle #20Continuity of useful action

Solution Approach 2:

The patent implements dynamics by adapting the segmentation model to dynamic video content. The recurrent module dynamically adjusts segmentation based on temporal context from previous frames, allowing the system to handle moving objects and changing scenes while maintaining temporal consistency, thus resolving the contradiction between speed and reliability.

Inventive Principle:
Principle #15Dynamics

2Reliability

If temporal information is integrated using recurrent modules, then temporal consistency is improved, but computational complexity increases

Engineering Contradiction:
Improvetemporal consistencyVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies segmentation by dividing the computational task into distinct modular components: a base segmentation network for spatial feature extraction and a recurrent module for temporal integration. This modular segmentation allows efficient computation at each stage while achieving overall temporal consistency, reducing the burden of computational complexity.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements partial action by applying recurrent temporal modeling selectively at key processing stages rather than throughout the entire network. The base segmentation network processes frames efficiently, and the recurrent module adds temporal consistency only where needed, avoiding excessive computation while maintaining reliability.

Inventive Principle:
Principle #16Partial or excessive action

3Measurement precision

If optical flow warping is applied to adapt internal states to pixel motion, then segmentation accuracy is improved, but processing time increases

Engineering Contradiction:
Improvesegmentation accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-computing optical flow fields and preparing warped feature representations in advance within the recurrent module. This preliminary preparation of motion-compensated features enables accurate segmentation without adding significant processing time during inference, as the computationally intensive warping operations are efficiently integrated into the recurrent processing pipeline.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3608844B1Methods for training a CRNN and for semantic segmentation of an inputted video using said crnn
Publication Date: 2025.07.02 NAVER CORP
  • EP3608844B1 patent drawingFigure 1a
  • EP3608844B1 patent drawingFigure 1b
  • EP3608844B1 patent drawingFigure 1c

AI summary

The present invention relates to a method for training a convolutional recurrent neural network, CRNN, for semantic segmentation in videos; the method being characterized in that it comprises the implementation, by a data processor (11a) of a first server (1a), of steps of: (a) Training from a base of training images already semantically segmented, a first convolutional neural network, CNN; (b) Training from a base of training videos already semantically segmented, a recurrent convolutional neural network, CRNN, corresponding to the first CNN wherein a convolutional layer has been replaced by a recurrent module having a hidden state; said training comprising, for each pair of successive frames (t - 1, t ∈ 1; T 2) of a video of said base of training videos already semantically segmented: (b1) Warping the internal state of the recurrent layer according to an estimated optical flow between the frames of the pair, so as to adapt the internal state to the motion of pixels between the frames of the pair; (b2) learning parameters of at least the recurrent module A method for semantic segmentation of an inputted video is further proposed.