UMAP Projections of Animals to 2D (duhaime.s3.amazonaws.com)

UMAP ZOO is a web demo that combines the UMAP dimensionality‑reduction algorithm with THREE.js to turn 3D Wavefront (.obj) animal models into interactive 2D projections. The site lets you visually explore how UMAP embeds complex 3D shape data into a flat map, making similarities, clusters and outliers among models immediately apparent. It’s a compact, hands‑on illustration of UMAP’s ability to preserve local neighborhood structure while revealing global relationships, useful for teaching, dataset exploration and quick sanity‑checks of shape collections. Technically, the demo imports 3D meshes (Wavefront files), converts mesh data into a feature representation (e.g., vertex coordinates, normals or other per‑mesh descriptors), then runs UMAP to compute 2D embeddings and displays them with THREE.js for interactive inspection. Important knobs include the choice of input features, distance metric, and UMAP hyperparameters (n_neighbors, min_dist), all of which strongly affect cluster tightness and global layout. The project highlights practical tradeoffs: UMAP is fast and preserves local topology but is stochastic and can distort global geometry, and preprocessing/scaling of mesh features matters. For researchers and practitioners, the site is a useful, intuitive tool to experiment with embedding choices and better understand how dimensionality reduction exposes shape relationships.