2D Gum-Net predicts unsupervised dense elastic deformation fields
Original title: 2D Gum-Net: Unsupervised Dense Elastic Image Registration via Symmetric Correlation and Spectral Pooling
Zero-shot on real fingerprint benchmarks, structural correlation rises by up to 15.7% on FVC2004 after training only on synthetic impressions.
Researchers present 2D Gum-Net, an end-to-end network that predicts dense elastic deformation fields without ground-truth labels. It adapts the 3D Geometric Unsupervised Matching Network for 2D registration, replacing the rigid head with a Dense Spatial Transformer and using DCT spectral pooling. On contact fingerprints, zero-shot evaluation on real benchmarks improves structural correlation and genuine match scores while imposter scores stay essentially unchanged.
Unconfirmed: Reported accuracy gains are the authors' own benchmark results and have not been independently confirmed.
- All three modules are redesigned in 2D from the original cryo-ET Gum-Net.
- DCT spectral pooling uses frozen-basis multiplications, about 50 times faster on GPU.
- A sparse control-point grid is lifted to full resolution by bicubic interpolation.
- Training uses only synthetic impressions; evaluation is zero-shot on real data.
- Genuine match scores rise; imposter scores remain essentially unchanged.
- AI models
- 2D Gum-Net
From the original article
1 min read at arxiv.orgDense non-rigid registration, the recovery of a pixel-wise deformation field that warps one image into geometric agreement with another, is fundamental across medical imaging, structural biology, and correspondence estimation, yet learning-based approaches face a severe bottleneck: ground-truth deformation fields are notoriously difficult to obtain for real data.
By Aarya Bhave · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.