6/6
Check out our project page: red-fairy.github.io/argus/
Paper: arxiv.org/abs/2504.07940

Joint with: Matt Wallingford, Ali Farhadi, Noah @snavely.bsky.social
, and Wei-Chiu Ma

5/6 But it doesn't stop there! Our model enables exciting potential applications like video stabilization without cropping, camera viewpoint control, and even interactive visual question answering.

4/6 Our model can predict scene dynamics: We capture a scene with a 360° camera and provide our model with a 60° FoV input of fixed viewing direction (left). The car's ground truth trajectory (middle) and our model's predicted trajectory (right) shows strong alignment.

3/6 Our solution? We've created a high-quality data pipeline leveraging vast online 360° videos, carefully curated to build effective training pairs.

Our model can generate spatially-coherent and temporally-consistent 360° videos from arbitrary in-the-wild perspective inputs.

2/6 ⬆️Input (left): Standard perspective video from a single viewpoint
Output (right): Our generated 360° panorama created from the input. The red box highlights the corresponding region from the original input.