Episode 3: 3D DepthAnything
3D AI Live/Replay/Episode 03
One photo in, one 3D point cloud out.
3D reconstruction usually needs many views. Here we do it from a single image: a vision transformer predicts the depth, and we project it into a coloured point cloud.
The brief
What we build in this session
You start with
A single image, with no other information.
You end with
A coloured point cloud (.ply).
How we get there
Depth Anything predicts a depth image, which we combine with the camera parameters to get 3D points.
The replay
Watch the full session
The materials
Everything you need to follow along
Go deeper
The papers and tools behind the session
- Vision Transformers, the paper→
- Depth Anything (CVPR 2024), the paper→
- Depth Anything, the model on Hugging Face→
- Depth Anything, the code on GitHub→
- CUDA Toolkit→
Where you are
You are in the free part of the academy
Four levels, each one a step deeper than the one before. Take them in your own order.
The live sessions
Free replays, open to everyone. This page is one of them.
The weekly tutorials
One hands-on 3D tutorial a week, free to read.
3The book
3D Data Science with Python: the full method, chapter by chapter.
4The programme
The 3D AI Architect: from 3D scan to deliverable, automated, with me beside you.
The next session is live, and your seat is free.
Save your seat once. You get every session, your personal link, the reminders and the replay.
