Optical Flow Localization for Detecting Novel Objects in Video

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing object localization systems fail to accurately identify 'novel' objects, which are objects not recognized by the known systems due to rarity, unusual appearance, or different visual characteristics.

Innovation Solution

A novel object localization system is trained using optical flow object localization data to identify objects based on movement, complementing known systems that rely on visual appearance, thereby generating training data to improve the accuracy of known systems.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a known object localization system is used to identify objects based on visual appearance, then objects with typical visual characteristics can be detected, but novel objects with unusual appearance or rarity cannot be accurately identified

Engineering Contradiction:
Improveobject localization accuracyVSAvoidability to detect novel objects
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent transitions from analyzing static visual appearance (2D image data) to incorporating temporal motion information (optical flow data across video frames). By adding the time dimension and motion characteristics, the system can detect novel objects that differ visually from training data but exhibit distinctive movement patterns, thereby resolving the contradiction between accuracy for typical objects and adaptability to novel objects

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

2Reliability

If an object localization system is trained on static images, then visual appearance features can be learned, but movement-based identification capabilities are lost

Engineering Contradiction:
Improvedetection consistencyVSAvoidmovement information
Core Design Contradiction:
ReliabilityVSLoss of information

Solution Approach 1:

The system pre-computes optical flow fields from video sequences before object detection. By preparing motion information in advance through optical flow calculation, the system preserves movement characteristics that would otherwise be lost, while maintaining the reliability of detection through consistent feature extraction across frames

Inventive Principle:
Principle #10Preliminary action

3Adaptability or versatility

If optical flow processing is applied to every video frame, then movement-based object detection can be achieved, but computational complexity increases

Engineering Contradiction:
Improvenovel object detection capabilityVSAvoidprocessing complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The system applies optical flow processing selectively rather than uniformly to all frames. By computing optical flow only where needed (e.g., in regions with detected motion or for specific novel object candidates), the system achieves novel object detection capability while avoiding the excessive computational burden of processing every frame completely

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3679520B1Processing images to localize novel objects
Publication Date: 2025.07.02 GOOGLE LLC
  • EP3679520B1 patent drawingFigure 1
  • EP3679520B1 patent drawingFigure 2
  • EP3679520B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training an optical flow object localization system and a novel object localization system. In a first aspect, the optical flow object localization system is trained to process an optical flow image to generate object localization data defining locations of objects depicted in a video frame corresponding to the optical flow image. In a second aspect, a novel object localization system is trained to process a video frame to generate object localization data defining locations of novel objects depicted in the video frame.