The sequence of images will be divided into two scenes

[xiazhenyu555]

When I was using the demo code, I found that the aerial images of the sequence would be split into two independent pose estimates. Have you ever encountered this situation?

[jashshah999]

The disconnected reconstruction problem happens because chunks are processed independently with no global constraints. A factor graph fixes this by enforcing overlap consistency + loop closure.

I built a tool that does exactly this:

1. Chunks with configurable overlap (default 3 frames)
2. Sim(3) alignment between overlapping frames
3. iSAM2 factor graph with DINOv2 loop closure
4. Exports to COLMAP / nerfstudio / PLY

At 300 frames it achieves 3.4x lower ATE than naive chunking.

https://github.com/jashshah999/vggt-factor-refinement

[xiazhenyu555]

The disconnected reconstruction problem happens because chunks are processed independently with no global constraints. A factor graph fixes this by enforcing overlap consistency + loop closure.

I built a tool that does exactly this:

1. Chunks with configurable overlap (default 3 frames)
2. Sim(3) alignment between overlapping frames
3. iSAM2 factor graph with DINOv2 loop closure
4. Exports to COLMAP / nerfstudio / PLY

At 300 frames it achieves 3.4x lower ATE than naive chunking.

https://github.com/jashshah999/vggt-factor-refinement

thanks! I will try this method!!

[jashshah999]

recon3d fixes disconnected chunks via Sim(3) overlap alignment + iSAM2 factor graph with DINOv2 loop closure. Tested on 300+ frame sequences. One command:

recon3d run aerial_video.mp4 --max-frames 300