Spatiotemporal Recycling Networks for Efficient Video Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current video/image processing techniques for semantic segmentation are inefficient and costly due to the repetitive computation of features in frames with high redundancy, leading to increased power consumption and processing costs.

Innovation Solution

Implementing a spatiotemporal recycling network that switches between a larger segmentation network and a lighter recycle network based on frame similarity, reusing features from previous frames to reduce computational load and enhance efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a larger segmentation network is used for every frame, then segmentation accuracy is improved, but processing time and power consumption increase

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

Solution Approach 1:

The system dynamically switches between two network configurations based on frame similarity assessment. When frames are highly similar, a lighter network is used; when differences are significant, the full larger network is deployed. This dynamic adaptation resolves the contradiction by adjusting processing intensity to actual needs rather than consistently using the maximum capability.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The system changes the operational parameters of the segmentation network by alternating between two distinct network configurations (full network vs. recycled feature network). This parameter change allows the system to maintain high accuracy when needed while reducing processing time during redundant frame sequences, thus resolving the accuracy-time tradeoff.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If a larger segmentation network is used for every frame, then segmentation accuracy is improved, but power consumption increases

Engineering Contradiction:
Improvesegmentation accuracyVSAvoidpower consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The system dynamically adjusts power consumption by switching between network configurations based on frame similarity. The lighter network consumes significantly less power while maintaining acceptable accuracy for similar frames, thus resolving the contradiction between accuracy and power usage.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

By changing the operational parameter from always-using-full-network to conditionally-switching-networks, the system achieves lower power consumption during redundant frame processing while preserving accuracy when frame content changes significantly.

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If feature computation is performed on every frame, then segmentation quality is maintained, but computational cost increases

Engineering Contradiction:
Improvesegmentation qualityVSAvoidprocessing efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The system performs preliminary assessment of frame similarity before deciding whether to execute full feature computation. This preliminary action allows the system to avoid unnecessary computational work on redundant frames while maintaining segmentation quality when it matters, thus resolving the contradiction between quality and efficiency.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system changes the computational parameter from constant full processing to variable processing intensity based on frame similarity metrics, achieving higher productivity without sacrificing segmentation quality in critical frames.

Inventive Principle:
Principle #35Parameter changes

4Productivity

If a lighter network is used for all frames, then processing efficiency is improved, but segmentation accuracy deteriorates

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidsegmentation accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The system dynamically selects network capacity based on actual frame content requirements. The lighter network is used during extended sequences of similar frames for efficiency, while the full network is activated when frame content changes significantly, ensuring accuracy is maintained when needed. This dynamic approach resolves the contradiction between efficiency and accuracy.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4200799B1Spatiotemporal recycling network
Publication Date: 2025.10.22 QUALCOMM INC
  • EP4200799B1 patent drawingFigure 1
  • EP4200799B1 patent drawingFigure 2
  • EP4200799B1 patent drawingFigure 3

AI summary

Systems, methods, and non-transitory media are provided for providing spatiotemporal recycling networks (e.g., for video segmentation). For example, a method can include obtaining video data including a current frame and one or more reference frames. The method can include determining, based on a comparison of the current frame and the one or more reference frames, a difference between the current frame and the one or more reference frames. Based on the difference being below a threshold, the method can include performing semantic segmentation of the current frame using a first neural network. The semantic segmentation can be performed based on higher-spatial resolution features extracted from the current frame by the first neural network and lower-resolution features extracted from the one or more reference frames by a second neural network. The first neural network has a smaller structure and/or a lower processing cost than the second neural network.