Paleo-Net — 3D Fossil Reconstruction

Overview

Led the development of a cutting-edge deep learning pipeline for the accurate 3D reconstruction of incomplete fossil specimens. Architected and trained a custom 3D U-NET model to process volumetric data, effectively predicting and restoring missing anatomical structures compromised by geological processes. This work bridges computational archaeology and paleontology, enabling researchers to digitally reconstruct specimens with unprecedented accuracy for morphological analysis and study.

Key Achievements

• Built PaleoVox, a Python library to procedurally generate training data featuring realistic deformations, fractures, and erosion on 3D bone voxels from the Ozboneviz database.
• Engineered a robust synthetic dataset by simulating realistic geological distortion including rotation, deformation, fractures, and erosion.
• Developed and trained a 3D Convolutional U-NET architecture on the synthetically augmented dataset.
• Implemented custom loss functions to accurately restore the original morphology of damaged specimens using GPU environments.
• Accepted for presentation at the XIX Congreso Nacional de Paleontología, recognizing the novel application of synthetic data in paleontological deep learning.