Lift4D: Harmonizing Single-View 3D Estimation for 4D Reconstruction In-the-Wild | Hacker News Reader

Lift4D: Harmonizing Single-View 3D Estimation for 4D Reconstruction In-the-Wild

lift4d.github.io

117 pointsilreb14 comments

Can someone use this on Elon's "my heart goes out to you" video, and compare it to... similar gestures. Would be interested to see exactly how close it is.

Looks cool. How’s it different from https://github.com/gaomingqi/sam-body4d ? My guess: Is sam-body4d focused on temporally consistent body reconstruction into an MHR model, and Lift4D creates the full scene and a model of the tracked object (that doesn’t need to be a human). I may be way off, but a comparison would help me!

I'm sorry but this just reminds me of this tng episode that inspired me as a teen https://www.youtube.com/watch?v=4Faiu360W7Q

Absolutely. Reminds me of the braindance sequences from Cyberpunk 2077

https://youtu.be/TV0ZUv4CMJ8?si=53etD-IUhsQkPpHt&t=1m41s

Just reading the first word of the title made me think about what a lift (elevator) simulator a la Elevator Saga[1] but in 4 spatial dimensions would look like.

[1] https://play.elevatorsaga.com

Really interesting, and a link to your GitHub with a "code coming soon". Please get us the tool already!

wondering how accurate the extrapolated distances are? Like the Rhino and the tree -- I wonder if the approximation could be useful in something like forensics from a security video?

> I wonder if the approximation could be useful in something like forensics from a security video?

No. It's just some plausible thing the AI made up. It's literally like asking an LLM to generate evidence of a crime based on the prompt you presented.

There are plenty of cool use cases, just not forensics.

Used properly I could imagine legitimate forensics uses to 3d reconstruction and splatting. For example as a "map" to find what underlying data (e.g. video frames) captured the thing.

This is insanely promising. What a time to be alive!

Two Minute papers fan? fun YT channel =3

Cool, so now when the swarm drones come after you, they can reconstruct the video on the fly from that single CCTV shot taken by flock, and pinpoint your location from the aggregated data by plantir!

Enemy of the State predicted this: https://www.youtube.com/watch?v=3EwZQddc3kY&t=11s

It's only a matter of time before it's in use by law enforcement: https://news.ycombinator.com/item?id=48520807

I hope no court ever accepts AI generated evidence.

Reply on news.ycombinator.com