Inspiration
Traditional re-enactment and reconstruction are slow, depend on real people and real-world logistics, and can make it difficult to capture what a witness thought, felt, or believed they saw.
What it does
Based on prompting, a user can prompt and reconstruct the original situtation from how they did it.
How we built it
Used google streat view api along with x2/h3 reactor to generate the live videos for real world objects, casting them into real world animationed objects with motion and 3d understanding of there mapping and the surrounding environment e.g. a person can hide behind a wall.
Challenges we ran into
World models are good at general generation, they are not good an more precise animation with specific coordinates and motion. Furthermore, it is harder to identify/specify the specific brand/look of the car
Accomplishments that we're proud of
Getting real world mappings and recognition to line up with google street view while using map box to understand what is going on in the surrounding area is to get depth of view and understand where objects are with respector.
What we learned
World models are very good, but only at specific more generalised videos, not more precise things like modelling/ more precise/exact motion. And learning to breakdown the problem effieicnet and testing fast and simply is good.
What's next for Reconstruction
Just a cool project idea to push world models to there limits for a real world specific use case.
Built With
- google-maps
- h3
- map-box
- next.js
- x2
Log in or sign up for Devpost to join the conversation.