Walk around a block, return to the starting point, and a video-derived 3D map may put that same corner in two different places. Small reconstruction errors accumulate.

HiP3R tackles the stitching problem. Existing AI models reconstruct overlapping video chunks; the new method first aligns their overall layout and scale, then reconciles conflicting camera positions for frames shared by several chunks. It works without retraining those underlying models.

The October 22 preprint reports tests across seven public benchmarks. The project page pairs real footage with reconstructions you can explore. Improved consistency still does not make a reconstruction a survey-grade measurement.

Explore the reconstructions · Paper and evaluation