The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) That's a VIBE: ML for Human Pose and Shape Estimation with Nikos Athanasiou, Muhammed Kocabas, Michael Black - #409
Sep 14, 2020
Join PhD students Nikos Athanasiou and Muhammed Kocabas, alongside Michael Black, the director of the Max Planck Institute for Intelligent Systems, as they unveil their groundbreaking VIBE research. They discuss the innovative adversarial learning framework for human pose and shape estimation and the significance of the AMASS dataset. The trio also dives into advancements in transforming sparse motion capture data into detailed 3D models and leveraging models with self-attention for enhanced accuracy in human motion understanding. A must-listen for AI enthusiasts!
AI Snips
Chapters
Transcript
Episode notes
3D Pose Estimation
- Human pose estimation can involve estimating key points or the whole body mesh.
- 3D pose estimation, focusing on body meshes, offers richer information than sparse key points.
Motion Capture from Video
- The Vibe method uses video sequences to estimate 3D pose and shape, capturing motion more smoothly.
- It leverages the continuity of motion in videos, unlike single-image methods.
Vibe's Performance
- The Vibe method outperforms a competing method by Anju Kanazawa, particularly in challenging videos.
- This improvement is achieved using a discriminator trained on the AMASS dataset to distinguish real from fake motions.