Learning the Depths of Moving People by Watching Frozen People

Explore a 22-minute Launchpad video that delves into the innovative technique of predicting depth for moving people by analyzing static images. Learn about the challenges of traditional stereo triangulation and how single-view depth prediction using multi-view supervision overcomes these limitations. Discover the process of transforming internet images into valuable training data and the application of the Mannequin Challenge to create a dataset for human depth prediction. Examine the progression from statues to people in depth estimation, and understand how the model is trained using RGB-only input. Witness the improvement in performance with increased input and the generation of pseudo-depth maps. Finally, explore the practical applications of this groundbreaking technology in various fields.

Intro
Goal
Where Could We Use This?
Existing Technologies
SLAM/MVS
Traditional Stereo Triangulation
Triangulation Here... Not So Good!
Single View Depth Prediction Using Multi-View Supervision
Internet Images Into Data
After Applying MV...
Depth Prediction on MegaDepth... A Lot Less Noisy!
Statues vs People
Mannequin Challenge
Training Data... Now On People
Get The Training Data
Train The Model (Using RGB- Only)
Prediction on Single RGB Image
More The Input Better The Performance
Getting Pseudo-Depth Map
Final Model Output
Application