PanSt3R: Multi-view Consistent Panoptic Segmentation
Published in International Conference on Computer Vision (ICCV), 2025
Lojze Zust, Yohann Cabon, Juliette Marrie, Leonid Antsfeld, Boris Chidlovskii, Jerome Revaud, Gabriela Csurka
PanSt3R is a unified approach for panoptic segmentation of 3D scenes from unposed 2D images. Prior methods typically rely on per-frame 2D panoptic predictions that are fused into an implicit 3D representation using costly test-time optimization. In contrast, PanSt3R jointly predicts 3D geometry and multi-view panoptic segmentation in a single forward pass, without requiring camera parameters.