Foreseeing the Invisible: Amodal Reconstruction of Leaf Fossil Images

Paleobotanists often work with fossil leaves that are partially obscured by rock, broken, or eroded, making it difficult to determine the original leaf shape—a critical feature for species identification and climate reconstruction. The authors address this by framing the recovery of missing leaf tissue as an amodal reconstruction task, aiming to predict the full silhouette and venation pattern from a single, incomplete RGB image.

The authors introduce AmodalDINO, a model that predicts four distinct masks: the visible leaf, the complete amodal leaf, the primary vein, and fine detail veins. Unlike existing amodal segmentation methods that require a visible mask as an input, AmodalDINO operates directly on RGB images. The authors identify two key design choices for success:

1. Full Fine-Tuning: Instead of freezing the DINOv3 backbone, they fine-tune the entire network at a low learning rate to allow the model to adapt to the specific geometry of leaf fossils.
2. Auxiliary Venation Heads: By adding heads that specifically predict primary and detail veins, the model is forced to learn a structural shape prior, as coherent venation can only exist within a complete, anatomically plausible leaf lamina.

The model was trained on a synthetic dataset generated using the NMNS Cleared Leaf Database, where complete leaves were projected onto randomized 3D stone models and then subjected to simulated breakage.

AmodalDINO achieves high accuracy on synthetic validation sets (95.0% Dice / 90.5% IoU) and demonstrates strong transferability to real-world fossil specimens. The authors show that the model is practical for real-world use, as it can be quantized to 4-bit weights to run entirely offline in a web browser. Furthermore, the model's architecture allows for the estimation of surface area through ruler-based calibration and supports generative visualization of living leaves.