README.md

metadata

title: 3D Model Search API
emoji: 🏢
colorFrom: blue
colorTo: indigo
sdk: docker
app_port: 7860
pinned: false
tags:
  - computer-vision
  - 3d
  - deep-learning
  - fastapi
  - pytorch
  - image-classification
  - similarity-search
  - pose-estimation
short_description: 3D building model classification and pose estimation
python_version: 3.12
app_file: main.py

Bachelor Project - 3D Model Search API

A FastAPI-based backend service for 3D building model classification, search, and pose estimation using deep learning models.

myapp - Main Application Server

The myapp folder contains the production-ready FastAPI server that provides REST API endpoints for 3D model search and classification.

Contents

• main.py - FastAPI application with CORS middleware and model loading lifecycle management
• routers/ - API route handlers:
  • buildings_router.py - Building model CRUD operations
  • buildings_search_router.py - Image-based 3D model search endpoints
• database/ - Database interaction layers:
  • buckets/ - B2 cloud storage operations
  • cache/ - Redis caching layer
  • metadata/ - Supabase metadata management
  • vectors/ - Pinecone vector database operations
• AR/ - AR model logic for pose estimation
• utils/ - Helper utilities including model loading functions
• trained_model/ - Trained neural network weights and model files
• models/ - Pydantic data models for request/response validation
• logging/ - Application logging configuration

Starting the Server with Docker Compose

The application uses Docker Compose with GPU support for running the FastAPI server and Redis cache.

Prerequisites

• Docker and Docker Compose installed
• NVIDIA GPU with Docker GPU support
• Environment variables configured in .env file

Required Environment Variables

Create a .env file in the project root with:

SUPABASE_URL=your_supabase_url
SUPABASE_KEY=your_supabase_key
B2_KEY_ID=your_b2_key_id
B2_APP_KEY=your_b2_app_key
B2_BUCKET_NAME=your_bucket_name
PINECONE_API_KEY=your_pinecone_api_key
PINECONE_REGION=your_pinecone_region
ENVIRONMENT=development

Start the Server

# Build and start all services (FastAPI server + Redis)
docker-compose up --build

The API will be available at http://localhost:8000

Stop the Server

docker-compose down

Technology Stack

• FastAPI - Modern Python web framework
• PyTorch - Deep learning framework for model inference
• Transformers - Hugging Face models for feature extraction
• Redis - In-memory caching
• Supabase - PostgreSQL database for metadata
• Pinecone - Vector database for similarity search
• B2 Cloud Storage - Object storage for 3D models and images
• Trimesh & PyRender - 3D model processing and rendering

Other Project Folders

classificator_training

Scripts and tools for training the 3D model classification neural network using prototype learning and feature extraction.

pose_estimation_training

Contains multiple approaches for camera pose estimation including YOLO-based pose detection, direct estimation methods, and feature matching algorithms.

blender

Blender Python scripts for automated 3D model rendering from different camera angles to generate training datasets for classification and pose estimation.

data_deploy_automation

Automation scripts for deploying 3D model data and metadata to cloud services through server in one script run (B2 storage, Supabase database, Pinecone vectors).

Development

The project uses Python 3.12 and includes dependency management via:

• requirements.txt - Production dependencies
• pyproject.toml - Poetry/UV package configuration