driverflow 0.3.0

AI-powered image annotation tool for Google Colab

DriverFlow — Intelligent Image Annotation Tool

A cloud-ready annotation tool powered by GroundingDINO for zero-shot object detection. Runs entirely in Google Colab — no local setup required.

Features

• Zero-shot detection — detect any objects using natural language prompts
• Interactive UI — upload images, adjust thresholds, preview results in real-time
• YOLO export — download annotations as a ZIP with annotations.txt + classes.txt
• Cloudflare Tunnel — optional persistent public URL with no account required

Quick Start (Google Colab)

!pip install driverflow

from driverflow import DriverFlow
DriverFlow().start()               # Colab proxy URL
# DriverFlow().start(tunnel=True)  # persistent public URL via Cloudflare

start() handles everything: installing GroundingDINO, downloading model weights, installing dependencies, and launching the web UI.

Usage

1. Upload an image — drag and drop or click to browse
2. Enter a text prompt — describe objects to detect, e.g. car . person . traffic light
3. Adjust thresholds (optional)
   • Box Threshold (default 0.35) — higher = only more confident detections
   • Text Threshold (default 0.25) — higher = stricter text-alignment matching
4. Click Detect — returns an annotated image and a summary table
5. Download — export as YOLO format ZIP

YOLO Export Format

annotations.txt — one detection per line:

<class_id> <cx> <cy> <w> <h>

All coordinates are normalized (0–1). class_id is 0-based, sorted alphabetically.

classes.txt — one class name per line, alphabetically sorted.

API

The tool runs a local FastAPI server with two endpoints:

POST /api/detect

Multipart form: image (file), text_prompt (str), box_threshold (float), text_threshold (float).

Returns JSON with detections, class_counts, annotated_image_b64, image_width, image_height.

POST /api/download_yolo

JSON body: detections, image_width, image_height.

Returns a ZIP file with annotations.txt and classes.txt.

Troubleshooting

Slow first run — GroundingDINO (~400 MB) and model weights are downloaded on first start(). Subsequent runs in the same Colab session are fast.

Low confidence / missing detections — lower the thresholds (e.g. 0.25 → 0.15) or use more specific prompts like red car instead of car.

Colab proxy expires — use tunnel=True for a persistent Cloudflare URL.

Model

GroundingDINO SwinT-OGC — Vision Transformer + BERT text encoder, ~400 MB checkpoint. Licensed under Apache 2.0 by IDEA-Research.

License

MIT — see LICENSE.