DriverFlow

DriverFlow is an AI-assisted image annotation and segmentation workspace for Google Colab. It combines Grounding DINO detection, SAM 2 segmentation, click refinement, versioned previews, and YOLO/raw-media export in a browser UI.

Quick Start

!pip install driverflow

from driverflow import DriverFlow

DriverFlow().start()              # Colab proxy URL
# DriverFlow().start(tunnel=True) # public Cloudflare quick-tunnel URL

DriverFlow starts the current workspace UI. The older single-image, detect-only UI is still available as DriverFlowOld for backwards compatibility.

from driverflow import DriverFlowOld

DriverFlowOld().start()

Workspace UI

The web UI lets you build a small in-memory workspace of images and videos:

• Import images, videos, or zip files by dragging onto the page, clicking the import button, or clicking the empty preview area.
• Import from the mocked DriverFlow Cloud tree. Cloud is not implemented yet; selected image files become 500x500 white PNGs and selected video files become plain 2-second MP4s.
• Use the left sidebar to browse raw, processed, and exported items with image/video filters.
• Click or drag one item into the preview area to make it the working item.
• Use the right tools panel to run image tools. Videos can be organized, previewed, and raw-exported, but no video tools are currently available.
• Scroll to the Data Workspace to view every version of the working item: raw media, detected boxes, segmented masks, and refined masks with clicks.
• Click or drag any Data Workspace card into the preview area to make that version authoritative for the next tool run or export.

Replacing a dirty working item warns before switching away from unexported derived work.

Models And Tools

Models are loaded on demand from the right-side model panel:

• Detect requires Grounding DINO and a raw image version.
• Segment requires SAM 2 and a detected image version with boxes.
• Refine requires SAM 2 and a segmented or refined image version with masks.

Trying to interact with a tool whose model is not loaded expands/highlights the required model. Heavy model dependencies and weights are installed lazily when models are loaded.

Export

The export button exports exactly the version currently shown in the preview:

• Raw image/video versions export the original media bytes.
• Detected image versions export YOLO bounding-box annotations.
• Segmented or refined image versions export YOLO segmentation annotations.

YOLO exports are ZIP files containing classes.txt and annotations.txt.

Library API

You can also use the pipeline directly from Python:

from driverflow import Pipeline, viz

pipe = Pipeline().setup(dino=True, sam=True)
det = pipe.detect("image.jpg", prompt="car . person . traffic light")
seg = pipe.segment(det)
viz.show_masks(seg)

Web API Shape

The workspace UI is backed by a local FastAPI server. The main endpoint groups are:

• POST /api/import/upload
• GET /api/import/cloud_tree
• POST /api/import/cloud_select
• GET /api/items
• GET /api/preview/{item_id}
• GET /api/preview/thumb/{item_id}
• GET /api/models/status
• POST /api/models/load_dino
• POST /api/models/load_sam
• GET /api/tools/registry
• POST /api/tools/{name}
• GET /api/export/{item_id}

The old /api/detect and /api/download_yolo endpoints belong to the legacy DriverFlowOld UI.

License

MIT - see LICENSE.