Selfie to Bust 3D Modeling Project Ideas

ISEF Category: Technology Enhances the Arts

Subcategory: 3D Modeling  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

One selfie can now become a 3D bust in minutes. That sounds like magic, but it runs on face landmarks, learned shape changes, and 3D printing. Your job is to ask a harder question, how close can a fast pipeline get to human likeness? That turns a cool demo into real research.

What Is It?

This project takes a single face photo and turns it into a printable 3D bust. The system first finds key points on the face, like the corners of the eyes, nose tip, and jawline. Then a learned decoder predicts how the face should bulge or sink in 3D, like a sculptor making tiny shape edits across a clay head.

A displacement map is like a height map for a surface. Bright and dark regions tell the printer or 3D model where to push out or pull in. MediaPipe FaceMesh gives the landmark data, and the decoder turns that data into a full face shape. Your science fair project can test whether that shape actually looks like the person, not just whether the software runs.

The key research idea is likeness. You can ask whether certain selfie conditions, face angles, or model settings produce busts that people rank as more realistic. That means your project sits at the intersection of computer vision, 3D modeling, and human perception.

Why This Is a Good Topic

This is a strong science fair topic because you can test it with clear inputs and clear outputs. You change one thing, like selfie angle, lighting, or decoder choice, and measure the effect on likeness scores or model quality. The topic also connects to practical problems in digital art, custom figurines, game avatars, and rapid prototyping. You can learn computer vision, 3D mesh editing, and experiment design without needing a chemistry or biology lab.

Research Questions

• How does selfie angle affect perceived likeness of the final bust?
• What is the effect of lighting consistency on landmark detection and bust accuracy?
• Does using more facial landmarks improve blind ranking scores for likeness?
• To what extent does the choice of displacement-map decoder change facial feature preservation?
• Which face regions, such as eyes, nose, or jawline, contribute most to perceived likeness?
• How does print resolution affect the visual similarity of the printed PLA bust?
• What is the effect of adding multiple selfies instead of one selfie on model quality?

Basic Materials

• A smartphone camera with fixed photo settings.
• A computer that can run Python and a face landmark model.
• MediaPipe FaceMesh installed in Python.
• Blender or another free 3D modeling tool.
• A 3D printer with PLA filament.
• A digital caliper for checking printed dimensions.
• A neutral backdrop and consistent desk lamp for photo capture.
• A small group of blind-rank judges.
• Consent forms for any human face photos you collect.

Advanced Materials

• A workstation with a dedicated GPU for model training.
• A dataset of face photos with paired 3D scans or dense depth labels.
• Open3D for mesh processing and comparison.
• PyTorch or TensorFlow for training the decoder.
• A structured-light or photogrammetry setup for ground-truth face scans.
• A high-resolution 3D printer with calibrated extrusion settings.
• MeshLab for mesh cleanup and alignment.
• ImageJ for measuring printed image or texture quality.

Software & Tools

• MediaPipe FaceMesh: Detects facial landmarks from selfies and gives you a structured face map.
• Blender: Lets you inspect, edit, and compare the generated 3D bust mesh.
• Python: Runs your pipeline, data logging, and statistical analysis.
• OpenCV: Handles image preprocessing, alignment, and quality checks.
• ImageJ: Measures image and texture features when you compare prints or renderings.

Experiment Steps

1. Define the likeness metric you will use, such as blind ranking, pairwise comparison, or a rubric.
2. Choose the one input variable you will change first, such as pose, lighting, or number of selfies.
3. Build a baseline pipeline that turns one photo into a mesh before you test any improvements.
4. Plan a comparison set, so each bust version can be judged against the same reference photo or scan.
5. Design controls that separate model quality from printing quality, since both can change the final result.
6. Select the statistical test that matches your data, then decide how you will report effect size and confidence.

Common Pitfalls

• Judges seeing the original selfie during scoring, which biases likeness ratings.
• Changing camera distance between photos, which makes face scale vary and confuses the model.
• Mixing print defects with model errors, which hides whether the decoder or the printer caused the problem.
• Training on too few faces, which makes the decoder memorize one face shape instead of general face structure.
• Using only one output metric, which can miss cases where the bust looks similar in one region but not another.

What Makes This Competitive

A stronger project would test more than one model setting and explain why one works better. You could compare landmark density, decoder architecture, or face pose correction with the same evaluation method. A competitive version would also separate technical accuracy from human perception, since those are not always the same. If you add careful controls, a clean ranking study, and a solid statistical analysis, the project starts to look like real research instead of a cool demo.

Project Variations

• Use full-face 3D scans as ground truth instead of blind rankings, then compare mesh similarity scores across model versions.
• Test whether busts made from profile photos, frontal photos, or averaged multi-angle inputs produce better likeness judgments.
• Compare printed PLA busts with digital renders only, so you can isolate whether printing artifacts change perceived realism.

Learn More

• MediaPipe documentation: Search the official MediaPipe site for FaceMesh examples, landmark maps, and pipeline details.
• Blender Manual: Use the free Blender documentation to learn mesh editing, scaling, and export settings.
• PyTorch Tutorials: Read the free official tutorials to learn how to train a small decoder model.
• OpenCV Documentation: Find image preprocessing and landmark alignment tools in the free OpenCV docs.
• PubMed: Search for review articles on facial perception, face recognition, and likeness judgment studies.
• Google Scholar: Search for peer-reviewed papers on 3D face reconstruction from a single image and displacement maps.