Deep Learning Digital Image Correlation
No subset size or step size — just load images and run.
GPU-accelerated dense field computation in under a second.
Native large deformation without initial guess.
Fully open source, runs on any NVIDIA GPU.
From raw speckle images to full-field strain maps in seconds.
End-to-end dense displacement and strain estimation.
Input image pairs
Deep optical flow
Dense u, v maps
εxx, εyy maps
RAFTcorr feeds speckle image pairs into a deep optical flow network (RAFT) to produce dense displacement fields. Strain fields are then computed via a virtual strain gauge method — no iterative optimization, no parameter tuning.
Validated across diverse experimental conditions.
Raw Image
von Mises Strain
Raw Image
von Mises Strain
Raw Image
von Mises Strain
Raw Image
Rotation
Honest comparison against established DIC tools.
| Metric | RAFTcorr | Ncorr | DICe | VIC-2D |
|---|---|---|---|---|
| Speed | Seconds | Minutes | Minutes | Minutes |
| Disp. Range | Large | Moderate | Moderate | Moderate |
| Parameters | Minimal | Multiple | Multiple | Multiple |
| Accuracy | 0.01–0.1 px | 0.01–0.1 px | 0.01–0.1 px | 0.01–0.1 px |
| Cost | Free | Free | Free | Commercial |
| GPU Required | Yes (NVIDIA) | No | No | No |
| GUI | ✓ | ✓ | ✓ | ✓ |
| 3D / Stereo | ✗ | ✗ | ✓ | ✓ |
| Post-Processing | Rich | Limited | Limited | Rich |
| Maintained | Active | Since 2020 | Since 2024 | Commercial |
Accuracy depends on image quality, speckle pattern, and calculation parameters for all methods. RAFTcorr’s key advantages are speed and large displacement range without parameter tuning.
The University of Texas at Austin
Lead Developer
Researcher
Researcher
Researcher
Principal Investigator
If you use RAFTcorr in your research, please cite our work.
@article{tong2026raftcorr,
title={RAFTcorr: An Adaptive, Open-Source, Deep Learning
Digital Image Correlation Framework for Dense
Displacement Measurement},
author={Tong, Zixiang and Bu, Lehu and Shi, Qihang
and Du, Runtian and Yang, Jin},
year={2026},
note={In preparation}
}