A simple multi-modal 3D asset retrieval system. This repo uses objaverse as a simple demonstration, while it can generalize and scale to any 3D dataset.

Demo built upon objaverse, using Cap3D 650k+ english captions, 650k+ translated captions ,and 260k+ gobjaverse asset renderings

• Text Search: Retrieve 3D assets in English or Chinese
• Image Search: Retrieve 3D assets using a single RGB image
• Cross-Modal Retrieval: Retrieve 3D assets using text2image or image2text similarities
• Dual Algorithms:
  • SigLip: Fast, English-only, per-image embeddings. Medium Retrieval Quality(WIP).
  • Qwen3-VL-Embedding: Bilingual, multi-image embeddings. High Retrieval Quality(Recommended).
• Vector Database: PostgreSQL with pgvector for efficient similarity search
• Web Interface: Beautiful Gradio UI with 3D model viewer
• REST API: FastAPI backend for programmatic access

┌─────────────────┐
│ English Captions│
└────────┬────────┘
         │
         ├──► Translation ──► Chinese Captions
         │
         ├──► SigLip Embeddings ──┐
         │    - Text (EN)          │
         │    - Images (per-view)  │
         │                         │
         └──► Qwen Embeddings ─────┤
              - Text (EN + CN)     │
              - Images (multi)     │
                                   │
                                   ▼
                         ┌──────────────────┐
                         │ PostgreSQL       │
                         │ + pgvector       │
                         └────────┬─────────┘
                                  │
                         ┌────────▼─────────┐
                         │  FastAPI Backend │
                         └────────┬─────────┘
                                  │
                         ┌────────▼─────────┐
                         │  Gradio Frontend │
                         └──────────────────┘

• Python 3.8+
• PostgreSQL 12+ with pgvector extension
• NVIDIA GPU (recommended for SigLip)
• DashScope API key (for Qwen)

1. Clone the repository

git clone https://github.com/3DSceneAgent/AssetRetrieval3D

2. Install dependencies:

   # optionally create new env 
   # conda create -n asset_retrieval python=3.10
   # conda activate asset_retrieval
   pip install -r requirements.txt

3. Set up environment variables:

   cp .env.example .env 
   # setup variables
   export DASHSCOPE_API_KEY="your-api-key"
   export DB_HOST="localhost"
   export DB_PORT="5432"
   export DB_USER="postgres"
   export DB_PASSWORD="your-password"

4. Configure PostgreSQL:

   • Install PostgreSQL
   • Install pgvector extension:

     CREATE EXTENSION vector;

Edit config.py to adjust settings:

• MAX_ASSETS: Set to a small number (e.g., 100) for testing
• Database credentials
• API keys
• Paths

Translate English captions to Chinese:

python scripts/01_translate_captions.py

This creates data/text_captions_cap3d_cn.json.

Note: This uses the Qwen batch API and may take hours for the full dataset.

python scripts/02_embed_siglip.py

This generates:

• Text embeddings (English only)
• Image embeddings (one per viewpoint)

python scripts/03_embed_qwen.py

This generates:

• Text embeddings (English and Chinese)
• Multi-image embeddings (8 images per asset)

Note: Can run both in parallel if you have resources.

python scripts/04_populate_database.py

This:

• Creates two databases: siglip_embeddings and qwen_embeddings
• Creates tables with pgvector columns
• Inserts all embeddings
• Creates vector indexes

python backend/app.py

Or with uvicorn:

uvicorn backend.app:app --host 0.0.0.0 --port 8000

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

In a new terminal:

python frontend/gradio_app.py

The UI will be available at http://localhost:7860

A comprehensive test client is provided to verify the backend service:

# Run all tests with default settings
python test_client.py

# Test only SigLip algorithm
python test_client.py --algorithm siglip

# Test only image search
python test_client.py --test-type image

# Run with verbose output (shows all results)
python test_client.py --verbose --num-tests 1

# Test against remote backend
python test_client.py --backend-url http://remote-server:8000

The test client will:

• Check backend health
• Test text search (English and Chinese for Qwen)
• Test image search using real assets from the database
• Test both uni-modal and cross-modal search
• Display results with similarity scores and captions /search/text

• MAX_ASSETS: Limit number of assets to process (for debugging)
• TRANSLATION_BATCH_SIZE: Captions per translation batch (default: 1000)
• EMBEDDING_BATCH_SIZE: Batch size for embedding generation (default: 100)
• QWEN_NUM_IMAGES: Number of images per asset for Qwen (default: 8)
• DEFAULT_TOP_K: Default number of search results (default: 10)

• Text → Text: Find assets with similar descriptions
• Image → Image: Find visually similar assets

• Text → Image: Find images matching text description
• Image → Text: Find text descriptions matching image

objaverse_retrieval/
├── config.py                 # Configuration
├── requirements.txt          # Dependencies
├── README.md                # This file
├── test_client.py           # Backend test client
│
├── data/                    # Data files
│   ├── text_captions_cap3d.json
│   ├── text_captions_cap3d_cn.json  (generated)
│   ├── gobjaverse/          # 3D asset images
│   └── gobjaverse_280k_index_to_objaverse.json
│
├── outputs/                 # Generated outputs
│   ├── embeddings/          # Saved embeddings
│   ├── translations/        # Translation results
│   └── batch_jsonl/         # API batch files
│
├── utils/                   # Utility modules
│   ├── data_loader.py
│   ├── image_utils.py
│   └── db_utils.py
│
├── scripts/                 # Processing scripts
│   ├── 01_translate_captions.py
│   ├── 02_embed_siglip.py
│   ├── 03_embed_qwen.py
│   └── 04_populate_database.py
│
├── backend/                 # FastAPI backend
│   ├── app.py
│   ├── embedding_service.py
│   └── vector_search.py
│
└── frontend/               # Gradio frontend
    └── gradio_app.py

The code and application is licensed under Apache2.0 License.

• objaverse
• objaverse_filter from kiui
• gobjaverse
• Cap3D