2026Systems

Agentic structure-from-motion

I am training a multimodal agent to call geometric tools (COLMAP, LoFTR, and retrieval) when a structure-from-motion pair is difficult. The research idea is Gabriele Berton's; this repository is my implementation, and it is still in progress.

Two photographs of a square with coloured lines marking matched points, shown at three saved training steps. — Fig. 1MegaDepth matches saved at training steps 10, 20, and 30. Three figures committed in Ryukijano/agentic-sfm. Not a screen recording of the agent.

Stack: SfM, COLMAP, LoFTR, Reinforcement Learning
Links: Code, Weights

Also

2026CUDA Blackwell Labs
2026PCOS edge agent
2024Dalton Mills VR Reconstruction
2023AWS AI/ML Scholar
B3tt3r: 3D Reconstruction