VisualGraph

Graph-based visual search and scene exploration using computer vision and Neo4j.

About

VisualGraph is a project for representing visual scenes as graphs. The initial idea is to detect objects in images, derive spatial relationships between them, and store this information in a Neo4j graph database.

For example, an image may be represented as:

Image
 ├── CONTAINS → Person
 ├── CONTAINS → Dog
 └── CONTAINS → Bicycle

Person ── NEAR ──→ Dog
Person ── LEFT_OF ──→ Bicycle

This graph representation can then be explored and queried using Cypher.

Initial Proposal

The first version of the project will focus on:

• object detection in images;
• extraction of bounding boxes and object classes;
• generation of spatial relationships between detected objects;
• representation of image scenes as property graphs;
• persistence and querying with Neo4j;
• visualization of the generated scene graph.

The initial relationships will be based on geometric information from bounding boxes, such as:

LEFT_OF
RIGHT_OF
ABOVE
BELOW
NEAR
OVERLAPS

Future versions may explore image embeddings, vector search, semantic relationships, and hybrid visual search.

Local development

Neo4j

Requirements: Docker with Compose, and uv for the backend.

cp .env.example .env        # then set NEO4J_PASSWORD (.env is not versioned)
docker compose up -d --wait # starts Neo4j and waits until it is healthy

• Neo4j Browser: http://localhost:7474 (Connect URL bolt://localhost:7687, user/password from .env)
• Data and logs are stored in ./neo4j/ (git-ignored). Stop with docker compose down; reset the database with docker compose down followed by deleting ./neo4j/data.

Verify the backend can connect (needs the container running):

cd backend
uv sync
uv run pytest -m integration

License

This project is licensed under the MIT License.