🏛️ Florence2 Plugin

Plugin Overview

This plugin allows you to apply Florence2 directly to your FiftyOne datasets.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone transformers<=4.49 accelerate einops timm torch 

Then, install the plugin:

fiftyone plugins download https://github.com/jacobmarks/fiftyone_florence2_plugin

You can also install requirements via:

fiftyone plugins requirements @jacbobmarks/florence2 --install

The Florence2 plugin integrates Microsoft's Florence2 Vision-Language Model with FiftyOne datasets, offering several powerful computer vision capabilities:

1. Caption Generation (CaptionWithFlorence2)

   • Generates image captions with different levels of detail
   • Customizable output field names
   • Supports both immediate and delegated execution

2. Object Detection (DetectWithFlorence2)

   • Multiple detection modes:
     • Standard object detection with default classes
     • Dense region captioning
     • Open vocabulary detection with custom text prompts
     • Region proposals
   • Flexible output field configuration

3. OCR (OCRWithFlorence2)

   • Performs text detection in images
   • Option to store region information (bounding boxes)
   • Configurable output fields

4. Caption-to-Phrase Grounding (CaptionToPhraseGroundingWithFlorence2)

   • Grounds specific phrases or captions by detecting relevant objects
   • Supports both direct text input and existing string fields from the dataset

5. Referring Expression Segmentation (ReferringExpressionSegmentationWithFlorence2)

   • Performs segmentation based on textual descriptions
   • Accepts either direct text input or references to existing dataset fields
   • Outputs segmentation masks for the described objects

All operators support:

• Custom model paths (defaults to "microsoft/Florence-2-base-ft")
• Delegated execution for resource-intensive tasks
• Flexible output field naming
• Integration with FiftyOne's dataset operations

ℹ👨🏽‍💻 Refer to the examples notebook for detailed examples for how to run each operator via the FiftyOne SDK!

Citation

@article{xiao2023florence,
  title={Florence-2: Advancing a unified representation for a variety of vision tasks},
  author={Xiao, Bin and Wu, Haiping and Xu, Weijian and Dai, Xiyang and Hu, Houdong and Lu, Yumao and Zeng, Michael and Liu, Ce and Yuan, Lu},
  journal={arXiv preprint arXiv:2311.06242},
  year={2023}
}