DiffPose
Differentiable projection matching for fast pose inference in cryo-EM
DiffPose is a differentiable framework for pose inference in cryo-electron microscopy. By replacing exhaustive projection matching with gradient-based optimization, DiffPose enables fast, accurate orientation refinement on large single-particle datasets.
Key ideas
- Differentiable projection operator that backpropagates pose gradients through the imaging model.
- GPU-efficient implementation suitable for hundreds of thousands of particles.
- Compatible with downstream 3D reconstruction pipelines.
Status: preprint available; software release in preparation.