Geometric Foundation Model
A model pretrained on large-scale 3D or multi-view data that outputs geometry-aware feature representations, in which three-dimensional priors such as depth, surface normals and spatial correspondence are implicitly encoded.
Timeline (last 90 days)
GFM features suffer from a viewpoint dependency problem: the same scene observed from different camera angles yields different extracted features, causing scene dynamics and camera motion to become entangled.
Mainstream 3D world models typically use GFM features directly as the world state, which causes the model to potentially learn feature changes caused by camera motion incorrectly as scene evolution.
Geometric foundation models are a class of models pretrained on large-scale 3D data or multi-view data that can output feature representations with geometric awareness, with representative works including pretrained backbones based on multi-view stereo, point clouds or neural radiance fields.
All Facts (3)
GFM features suffer from a viewpoint dependency problem: the same scene observed from different camera angles yields different extracted features, causing scene dynamics and camera motion to become entangled.
50%UnverifiedMainstream 3D world models typically use GFM features directly as the world state, which causes the model to potentially learn feature changes caused by camera motion incorrectly as scene evolution.
50%UnverifiedGeometric foundation models are a class of models pretrained on large-scale 3D data or multi-view data that can output feature representations with geometric awareness, with representative works including pretrained backbones based on multi-view stereo, point clouds or neural radiance fields.
50%