Research framework for globally consistent 3D reconstruction from long RGB videos using chunked priors and test-time adaptation.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
GeoWeaver is a research method for reconstructing camera trajectories and point clouds from long RGB video sequences. Feed-forward reconstruction models predict good depth and pose but cannot process long sequences jointly because of memory cost. GeoWeaver splits a sequence into short chunks that share one frame, and a Geometric Prior Model (GPM) predicts depth, confidence, camera pose and intrinsics for each chunk.
A Test-Time Adaptation (TTA) stage then assembles the chunks. It does sequential initialization, global chunk-level Sim(3) alignment, and coarse-to-fine refinement of camera poses, affine depth corrections and intrinsics. Dense correspondences supply adjacent, cross-chunk and long-range constraints. A robust objective jointly optimizes 2D reprojection and 3D consistency residuals. The authors report that the same TTA procedure also improves trajectory estimates of other geometric prior models.
The project page shows reconstruction videos (Courthouse, Barn, Meeting Room, Virtual KITTI 2, Keble College, Blenheim) and a per-scene comparison on Oxford Spires against methods such as DA3, VGGT-Long, Scal3R, LoGeR, LingBot-Map and AMB3R. At the time of the provided material only the paper (arXiv preprint, 2026) and project page are available. No code, weights or demo were found, so it is not yet usable by end users.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Reconstructs long videos into consistent 3D scenes by chunking, predicting depth and camera pose, then stitching chunks to reduce drift. Lowest average error on benchmarks; only a paper is released so far.
Closest neighbours by meaning, not just by category.
Research
Post-training method that refines feed-forward 3D reconstruction models (Pi3, DA3, VGGT) on unlabeled video.
Model
Research framework that turns posed images with instance masks and 3D boxes into per-object meshes for cluttered scenes.
Model
Research autoencoder with a geometry-native latent space for 3D-consistent video and world generation.
Model
Research model that predicts layered, pixel-aligned 3D point clouds, including hidden geometry, from one image or video.
Research
Research method that reconstructs complete 4D (3D plus time) scenes from a single monocular video using Gaussian Splatting.
Research · Preferred Networks
Generative method from Preferred Networks that reconstructs high-resolution 3D scene meshes from posed multi-view images.
