Fabio Poiesi

Relative 6D poses between identical unknown objects can now be resolved from a single image with zero CAD models, reference templates, or task-specific training.

Ditch the heavy 3D encoders: Fase3D uses Fourier transforms to build a surprisingly efficient and scalable 3D large multimodal model directly from point clouds.

By blending geometric and appearance cues within a conditional flow matching framework, Flose cracks the long-standing problem of 6D pose estimation for symmetrical objects.