Semantic Video Object Segmentation via Spatio-Temporal Graph Inference

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for semantic video object segmentation fail to capture long-range and high-level contexts due to changing object appearance and occlusions, leading to significant errors in object detection and segmentation.

Innovation Solution

A novel graphical model is constructed using tracklets, object proposals, and superpixels to integrate multiple scale contextual information, with weights calculated for connections between nodes to facilitate joint inference of superpixel and object proposal likelihoods, enabling accurate semantic object segmentation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If object detection is used to build explicit semantic notion of video objects, then middle- and high-level visual information is incorporated, but long-range and high-level contexts are not captured leading to significant errors

Engineering Contradiction:
Improveobject detection accuracyVSAvoidlong-range context information
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The patent extends the analysis from single-frame object detection to multi-frame temporal sequences by introducing tracklets and temporal graph structures. This dimensional extension from spatial to spatio-temporal domain enables capture of long-range contexts while maintaining object detection precision through temporal consistency constraints.

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

Solution Approach 2:

The patent introduces tracklets as intermediary structures that bridge individual object detections across multiple frames. These tracklets serve as mediators that aggregate temporal information and propagate contextual cues, enabling the system to capture long-range dependencies while maintaining accurate object segmentation through the temporal graph inference mechanism.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Extent of automation

If semantic video object segmentation is performed without human intervention, then automation is improved, but accuracy decreases due to changing object appearance and occlusions

Engineering Contradiction:
Improveautomatic segmentationVSAvoidsegmentation accuracy
Core Design Contradiction:
Extent of automationVSMeasurement precision

Solution Approach 1:

The patent implements feedback mechanisms through the temporal graph model where segmentation results from previous frames inform current frame analysis. The graph inference process continuously refines object segmentations by propagating contextual constraints backward and forward in time, enabling fully automatic segmentation to maintain high accuracy despite appearance changes and occlusions.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent performs preliminary action by pre-processing video frames to generate tracklets and object proposals before final segmentation. This preliminary temporal association and proposal generation creates a structured foundation that guides the subsequent graph inference, enabling fully automatic processing to achieve accurate segmentation without human intervention.

Inventive Principle:
Principle #10Preliminary action

3Device complexity

If a simple object detection approach is used, then device complexity is reduced, but segmentation accuracy deteriorates due to inability to capture temporal contexts

Engineering Contradiction:
Improveprocessing system complexityVSAvoidsegmentation accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent segments the complex segmentation task into distinct components: tracklet generation, object proposal extraction, superpixel segmentation, and graph inference. This modular segmentation of the processing pipeline manages system complexity by breaking down the temporal context capture into manageable stages while maintaining high segmentation accuracy through the coordinated interaction of these components.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3447727B1A method, an apparatus and a computer program product for object detection
Publication Date: 2021.01.13 NOKIA TECHNOLOGIES OY
  • EP3447727B1 patent drawingFigure 1~2
  • EP3447727B1 patent drawingFigure 3
  • EP3447727B1 patent drawingFigure 4~5

AI summary

A method, comprising: receiving a video comprising video frames as an input; generating set of object proposals from the video; generating object tracklets comprising regions appearing in consecutive frames of the video, said regions corresponding to object proposals with a predetermined level of confidence; constructing a graph comprising the object proposals and superpixels grouped from the pixels of the frame; calculating a first cost function for superpixel likelihoods in terms of the object proposals in an object tracklet; calculating a second cost function for object proposal likelihoods in terms of the superpixels in the frame; minimizing the first and the second cost function with each other; computing posterior probabilities of each superpixel; and assigning each superpixel with an object class with a maximum posterior probability to constitute semantic object segmentation.