Temporal Representation Learning on Monocular Videos for 3D Human Pose Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Extracting temporal information from monocular videos is challenging.
• Existing contrastive self-supervised learning (CSS) methods often treat temporal features simplistically.
• Accurate human pose estimation requires robust temporal feature extraction.

Purpose of the Study:

• To propose an unsupervised feature extraction method for capturing temporal information in monocular videos.
• To improve human pose estimation using enhanced temporal feature extraction.
• To develop a novel CSS approach that disentangles time-variant and time-invariant features.

Main Methods:

• Detecting and encoding subjects of interest in each video frame.
• Leveraging contrastive self-supervised learning (CSS) with disentangled latent vectors.
• Applying contrastive loss to time-variant features with gradual transition objectives.
• Incorporating input reconstruction for enhanced feature learning.

Main Results:

• Reduced human pose estimation error by approximately 50% compared to standard CSS methods.
• Outperformed other unsupervised single-view methods for temporal feature extraction.
• Achieved performance comparable to multi-view techniques in pose estimation.
• Significantly improved 3D pose estimation accuracy when 2D pose information was available.

Conclusions:

• The proposed method effectively captures rich temporal features from monocular videos.
• Disentangling time-variant and time-invariant components enhances feature representation.
• This approach offers a powerful unsupervised solution for human pose estimation and related tasks.