Permutation-equivariant feed-forward 3D reconstruction from a set of [Image] views with Pi3. One reference-free joint pass regresses per-view [CameraPose] and [PointMap3D], or a single fused PointCloud, with no privileged view ordering.
Typical backends
Multi-view point cloud. Load a set of overlapping stills, fuse them into one cloud, and pass it to a downstream 3D consumer.
Trajectory and structure from video. Decode frames from a sweep and publish the recovered per-view poses and point maps.
Caveats
- Hard constraintAt least two overlapping views of the same scene are needed per tick; the model has no privileged reference view, but a single view gives a degenerate pose and non-overlapping views reconstruct poorly.
- AccuracyGeometry is scale-invariant rather than metric, and no camera intrinsics are regressed — only poses, world point maps, and the fused cloud — so calibrate against a known distance when absolute scale matters.
- I/O contractPer-view outputs are produced at a shared downscaled grid bounded by a per-view pixel budget, not the source resolution; the confidence floor (a 0-to-1 probability) and the point cap apply only to the fused PointCloud arm.
- CostReconstruction is GPU-heavy and memory scales with the view count and the per-view pixel budget, which is the main lever for large images; the per-tick view count is capped to bound peak memory.
- State lifetimeThe model and weights are loaded once at startup, downloading from the model hub on a cold container, so the first batch needs outbound network.