Overview
- Researchers at Helmholtz-Zentrum Berlin and the University of Edinburgh report their method in PNAS and release the DinoTracker app for public use.
- The network was trained on nearly 2,000 real tracks plus millions of augmented variants and discovered eight key axes of variation, including toe spread and heel position.
- Only footprint outlines were input during training, reducing human-labeling bias and yielding about 80–93% concordance with expert classifications.
- Model outputs highlight hypotheses that some tracks over 200 million years old show bird-like features and that Isle of Skye prints cluster with early ornithopods, pending further evidence.
- DinoTracker is available via GitHub, enabling users to upload photos or sketches of tracks for instant similarity analysis to support research and citizen science.