Unsupervised Image Keypoint Learning Through Temporal Transport

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing machine learning models struggle to accurately identify keypoint locations in images without manual supervision, which hinders effective interaction and control of agents like robots and autonomous vehicles in dynamic environments.

Innovation Solution

A system and method for training a keypoint extraction machine learning model using spatio-temporal and temporal transport techniques, which processes input images to generate accurate keypoint locations on objects, enabling improved control and exploration of environments.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If unsupervised learning is used to train keypoint extraction models, then manual annotation effort is reduced, but keypoint location accuracy deteriorates

Engineering Contradiction:
Improvetraining effortVSAvoidkeypoint location accuracy
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The patent applies dynamics by making the training process adaptive through temporal transport. The model dynamically adjusts its learning based on temporal relationships between consecutive video frames, allowing the keypoint detection to evolve and improve accuracy over time without manual intervention. This dynamic approach resolves the contradiction by enabling the system to self-improve accuracy through temporal context rather than requiring static manual annotations.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent introduces temporal transport as an intermediary mechanism that bridges the gap between unsupervised learning and accurate keypoint detection. By using optical flow and temporal relationships as intermediaries, the system transfers information between frames to guide keypoint localization, effectively mediating between the lack of manual supervision and the need for precise keypoint locations.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If traditional supervised learning is used for keypoint detection, then keypoint location accuracy is improved, but training time and computational resources increase

Engineering Contradiction:
Improvekeypoint location accuracyVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-computing temporal transport features and optical flow information during the training phase. These pre-computed temporal representations are then reused during inference, eliminating the need for time-consuming manual annotation while maintaining accuracy. The temporal context is prepared in advance, allowing fast real-time keypoint detection without sacrificing precision.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses copying by replicating temporal patterns and optical flow information across multiple frames. Instead of manually annotating each frame individually (which would be time-consuming), the system copies temporal relationships from reference frames to target frames, significantly reducing training time while preserving keypoint location accuracy through the copied temporal context.

Inventive Principle:
Principle #26Copying

3Speed

If static image processing is used for keypoint detection, then processing speed is maintained, but ability to capture temporal dynamics deteriorates

Engineering Contradiction:
Improveprocessing speedVSAvoidtemporal dynamics capture
Core Design Contradiction:
SpeedVSAdaptability or versatility

Solution Approach 1:

The patent applies dynamics by integrating temporal transport mechanisms that explicitly model motion and change over time. The system processes video sequences by leveraging temporal relationships between frames, allowing it to capture dynamic keypoint movements while maintaining processing speed through efficient temporal feature extraction and reuse across the sequence.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent ensures continuity of useful action by maintaining temporal coherence across the entire video sequence. The temporal transport mechanism continuously leverages information from previous frames to inform current keypoint detection, creating an unbroken chain of useful temporal information that enhances adaptability to dynamic scenes without requiring reprocessing of individual frames, thus preserving processing speed.

Inventive Principle:
Principle #20Continuity of useful action

Data Source

PatentEP3935566B1Unsupervised learning of object keypoint locations in images through temporal transport or spatio-temporal transport
Publication Date: 2025.09.10 GDM HOLDING LLC
  • EP3935566B1 patent drawingFigure 1
  • EP3935566B1 patent drawingFigure 2
  • EP3935566B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for unsupervised learning of object keypoint locations in images. In particular, a keypoint extraction machine learning model having a plurality of keypoint model parameters is trained to receive an input image and to process the input image in accordance with the keypoint model parameters to generate a plurality of keypoint locations in the input image. The machine learning model is trained using either temporal transport or spatio-temporal transport.