4D-LRM: Large Space-Time Reconstruction Model From and To Any View at Any Time

Summary

This paper talks about 4D-LRM, a large model that can reconstruct objects in both space and time from a few images taken at different angles and moments, allowing it to generate views from any position and time.

What's the problem?

The problem is that existing methods either need many images, are slow, or don't handle movement and time well, making it hard to generate smooth, accurate 4D scenes.

What's the solution?

The researchers developed a Transformer-based model that learns a unified space-time representation using 4D Gaussian primitives, which lets the model predict detailed geometry and appearance quickly and accurately from sparse inputs, supporting any new viewpoint and time.

Why it matters?

This matters because it advances the ability to create dynamic 3D models and animations efficiently, which can be used in gaming, movies, virtual reality, and other fields where realistic and flexible visual scenes are needed.