Temporal Consistency Evaluation for Semantic Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for evaluating semantic image segmentation in video sequences fail to capture temporal characteristics and do not allow for real-time analysis, making them inadequate for assessing the quality and stability of deep neural networks in applications like highly automated driving.

Innovation Solution

A method that involves receiving sequential image frames, determining semantic segmentation predictions, estimating displacements between frames, applying these displacements to generate expected predictions, and evaluating the consistency of these predictions using optical flow to calculate temporal consistency metrics, enabling real-time assessment of semantic image segmentation stability.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If mIoU metric is used to evaluate semantic segmentation quality, then segmentation accuracy is measured, but temporal characteristics of video sequences are not captured

Engineering Contradiction:
Improvesegmentation accuracy measurementVSAvoidtemporal characteristics information
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The patent transitions from evaluating segmentation only in the spatial dimension (mIoU on individual frames) to incorporating the temporal dimension by comparing segmentation predictions across multiple sequential frames. This is achieved by computing temporal consistency metrics that measure the stability of segmentation labels over time, thereby capturing temporal characteristics that were previously lost.

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

2Measurement precision

If consistency is calculated over complete video sequence using large displacement optical flow, then temporal consistency is measured, but real-time usage is not possible

Engineering Contradiction:
Improvetemporal consistency measurementVSAvoidreal-time processing capability
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent divides the video sequence into overlapping segments or windows of multiple consecutive frames rather than processing the entire sequence at once. This allows temporal consistency to be computed locally for each segment, enabling real-time processing while maintaining measurement precision. The segmentation of the temporal dimension allows the system to provide timely feedback without the computational burden of analyzing complete long sequences.

Inventive Principle:
Principle #1Segmentation

3Measurement precision

If ground truth labels are obtained through hand-labeling, then segmentation quality evaluation is possible, but evaluation cost increases

Engineering Contradiction:
Improvesegmentation quality evaluation accuracyVSAvoidevaluation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent uses the semantic segmentation predictions from the neural network itself as proxies for ground truth labels. By comparing predictions across multiple frames and measuring their temporal consistency, the system creates a self-evaluation mechanism that does not require expensive manual annotation. This copying approach allows the model to evaluate its own temporal stability using its own outputs.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP3896651A1Method and apparatus for evaluating temporal characteristics of semantic image segmentation
Publication Date: 2021.10.20 VOLKSWAGEN AG
  • EP3896651A1 patent drawingFigure 1
  • EP3896651A1 patent drawingFigure 2
  • EP3896651A1 patent drawingFigure 3

AI summary

The present invention is related to a method, a computer program, and an apparatus for evaluating temporal characteristics of semantic image segmentation, in particular for detecting perception instabilities on temporally sequential images of a video sequence. The invention is further related to an in-car control unit or a backend system, which make use of such a method or apparatus, and to a vehicle comprising such an in-car control unit. In a first step, image data of two sequential image frames are received (10). Then semantic segmentation predictions for each of the two sequential image frames are determined (11). Displacements between the image data of the two sequential image frames are estimated (12). The estimated displacements are applied (13) to the semantic segmentation prediction of the first of the two sequential image frames to generate an expected semantic segmentation prediction for the second of the two sequential image frames. The semantic segmentation prediction of the second of the two sequential image frames is evaluated (14) based on the expected semantic segmentation prediction for the second of the two sequential image frames.