3D Hand Pose Estimation Using Temporal Self-Supervision

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Estimating 3D hand poses from RGB images is challenging due to the difficulty in annotating 3D poses, with limited available datasets and the need for explicit 3D annotations, which complicates training deep models.

Innovation Solution

A self-supervised learning model, called temporal-aware self-supervised network (TASSN), uses temporal consistency constraints to estimate 3D hand poses and meshes from videos annotated with only 2D key frame positions, leveraging optical flow and graph convolutional networks to infer 3D hand meshes without explicit 3D annotations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If explicit 3D annotations are used for training deep models, then measurement precision of hand pose estimation is improved, but device complexity and data annotation difficulty increase significantly

Engineering Contradiction:
Improvehand pose estimation accuracyVSAvoidannotation complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system performs self-supervised learning by automatically generating supervision signals from optical flow and temporal consistency constraints without requiring manual 3D annotations. The model learns to estimate 3D hand poses by exploiting temporal relationships in video sequences, where the system serves itself by creating its own training labels from readily available 2D video data.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

Optical flow serves as an intermediary that bridges 2D video frames to provide implicit 3D motion information. By computing optical flow between consecutive frames and enforcing temporal consistency, the system indirectly obtains supervision signals that guide 3D pose estimation without direct 3D annotation input.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If 3D annotated datasets are collected for training, then estimation accuracy is improved, but loss of time and productivity decrease due to manual annotation requirements

Engineering Contradiction:
Improve3D hand pose estimation accuracyVSAvoidannotation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system eliminates manual annotation time by performing self-supervised learning. Training data is automatically generated from unannotated video sequences through optical flow computation and temporal consistency enforcement, allowing the model to learn 3D pose estimation without human intervention in the annotation process.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system performs preliminary computation of optical flow and temporal consistency constraints on raw video data before actual pose estimation. This preliminary processing creates ready-to-use supervision signals that can be directly applied to train the model, eliminating the need for time-consuming manual 3D annotation during data preparation.

Inventive Principle:
Principle #10Preliminary action

3Reliability

If temporal consistency constraints are enforced, then reliability of 3D pose estimation is improved, but device complexity increases due to additional computational requirements

Engineering Contradiction:
Improve3D pose estimation reliabilityVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The system enforces temporal consistency by comparing predicted 3D poses across consecutive frames with optical flow-derived motion constraints. This feedback mechanism adjusts the pose estimates to ensure temporal coherence, improving reliability by preventing temporally inconsistent predictions while using computationally efficient optical flow comparisons.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP4094249B1Video-based 3D hand pose and mesh estimation based on temporal-aware self-supervised learning
Publication Date: 2025.11.12 TENCENT AMERICA LLC
  • EP4094249B1 patent drawingFigure 1
  • EP4094249B1 patent drawingFigure 2
  • EP4094249B1 patent drawingFigure 3

AI summary

A method, computer program, and computer system is provided for estimating three-dimensional hand poses in images. Data corresponding to two hand images is receive, and an optical flow value corresponding to a change in a hand gesture in the received hand image data is calculate. A heat map is generated based on the calculated optical flow, and a hand mesh map is estimated based on the generated heat map. A hand pose present within the hand images is determined based on the estimated hand mesh map.