I am currently working on 3D Gaussian Splatting (3DGS) for indoor room reconstruction. My pipeline uses VGGT-1B, followed by Bundle Adjustment (BA), to supply camera poses, images, and point clouds (derived from depth and poses) to 3DGS.
When processing the full sequence of 470+ images at once, the VGGT confidence drops sharply. However, running inference on smaller subsets yields good results. To scale this up, I tried chunking the dataset into segments of 60 images with a 30-image overlap, running inference per chunk, merging the scenes, and applying Global Adjustment. Unfortunately, the final reconstruction results remain poor.
Can any help with this
One likely culprit is the merge, not 3DGS: each chunk’s VGGT poses live in that chunk’s own coordinate frame and scale, so merged clouds can put the same wall in several places at once, and global adjustment starting from there rarely recovers. Align neighbouring chunks first — fit a similarity transform (rotation, translation, and scale) on the 30 shared images’ camera positions, apply it before merging, and check the overlap residual: if shared cameras still disagree after alignment, the chunk poses are the problem; if they agree, then run your joint adjustment. That one number tells you whether to fix alignment or look downstream.