ReLoop-3D

A minimal pipeline to classify rendered views of simple 3D meshes. Includes a Gradio app for uploading a mesh, rendering multiple views, and running a trained classifier.

Requirements

• Windows 10/11 or macOS/Linux
• Python 3.10+
• GPU optional (PyTorch supports CPU)

Install

python -m venv .venv
# Windows PowerShell
. .venv/Scripts/Activate.ps1
# macOS/Linux
# source .venv/bin/activate

pip install -r requirements.txt

Run the app

python -m src.app

What you get:

• Image tab: test the trained model on a single image.
• Image → 3D tab: convert image(s) to 3D mesh.
  • Single mode: depth estimation (MiDaS) → mesh (fast).
  • Multi-view mode: Open3D point cloud fusion → mesh (better quality, no external deps).
• 3D Model tab: upload .obj/.ply/.stl/.glb/.gltf, see live preview, render N views, classify.

Train a model (toy example)

This repo includes simple classifiers (CNN or ViT-tiny) trained on renderings. Adjust to your data as needed.

# CNN (image size 128)
python -m src.train --data_dir data --out_dir outputs --epochs 5 --model cnn

# ViT-tiny (image size 224)
python -m src.train --data_dir data --out_dir outputs --epochs 5 --model vit_tiny

Active Learning

Use the Active Learning tab in the UI:

1. Score Unlabeled: point to mesh folder, choose acquisition (entropy/margin), click Score.
2. Edit labels in the table.
3. Apply Labels: renders meshes into data/images/<label>/ and logs to annotations/labeled.csv.
4. Retrain with the updated dataset.

References

• Open3D: https://github.com/isl-org/Open3D

Citation

If you use this work, please cite:

@software{ReLoop3D,
  title={ReLoop3D: 3D Active Learning with Feedback for Machine Learning},
  author={Abaryan},
  year={2025},
  url={https://github.com/r-abaryan/ReLoop-3D}
}