Project Description:
This project addresses a critical bottleneck in entomological research: the time-intensive process of preparing 3D-scanned ant specimens for analysis, animation, and visualization. Evolutionary biologist Dr. Evan Economo has countless high-resolution 3D models of ants, generated from micro-CT scans. Due to the crumpling of specimens in preservation and storage, however, these models are difficult to study and excessively labor-intensive to rig manually – often requiring months of work per specimen.
To streamline this workflow, we leveraged Blender scripting and computer vision machine learning models to develop an automated rigging pipeline. Our systems preprocesses 3D ant models, generates synthetic training data using a custom Blender-based rendering script, and then trains a pose estimation model (DeepLabCut) to identify key joint positions from multi-view images.
When a user uploads a new 3D ant model, our pipeline captures images of it from multiple angles, applies the trained pose estimation model, and then maps the predicted joint locations back into 3D space. This enables us to automatically fit a base armature to the model, resulting in a fully articulable, rigged ant model suitable for scientific analysis, VR/AR environments, animation, and other multimedia or machine learning applications.
