🔍🤖 Vision Agent

Vision Agent is a library that helps you utilize agent frameworks for your vision tasks. Many current vision problems can easily take hours or days to solve, you need to find the right model, figure out how to use it, possibly write programming logic around it to accomplish the task you want or even more expensive, train your own model. Vision Agent aims to provide an in-seconds experience by allowing users to describe their problem in text and utilizing agent frameworks to solve the task for them. Check out our discord for updates and roadmaps!

Documentation

Vision Agent Library Docs

Getting Started

Installation

To get started, you can install the library using pip:

pip install vision-agent

Ensure you have an OpenAI API key and set it as an environment variable (if you are using Azure OpenAI please see the Azure setup section):

export OPENAI_API_KEY="your-api-key"

Vision Agents

You can interact with the agents as you would with any LLM or LMM model:

>>> from vision_agent.agent import VisionAgent
>>> agent = VisionAgent()
>>> agent("What percentage of the area of this jar is filled with coffee beans?", image="jar.jpg")
"The percentage of area of the jar filled with coffee beans is 25%."

To better understand how the model came up with it's answer, you can also run it in debug mode by passing in the verbose argument:

>>> agent = VisionAgent(verbose=True)

You can also have it return the workflow it used to complete the task along with all the individual steps and tools to get the answer:

>>> resp, workflow = agent.chat_with_workflow([{"role": "user", "content": "What percentage of the area of this jar is filled with coffee beans?"}], image="jar.jpg")
>>> print(workflow)
[{"task": "Segment the jar using 'grounding_sam_'.",
  "tool": "grounding_sam_",
  "parameters": {"prompt": "jar", "image": "jar.jpg"},
  "call_results": [[
    {
      "labels": ["jar"],
      "scores": [0.99],
      "bboxes": [
        [0.58, 0.2, 0.72, 0.45],
      ],
      "masks": "mask.png"
    }
  ]],
  "answer": "The jar is located at [0.58, 0.2, 0.72, 0.45].",
},
{"visualize_output": "final_output.png"}]

You can also provide reference data for the model to utilize. For example, if you want to utilize VisualPromptCounting:

agent(
    "How many apples are in this image?",
    image="apples.jpg",
    reference_data={"bbox": [0.1, 0.11, 0.24, 0.25]},
)

Where [0.1, 0.11, 0.24, 0.25] is the normalized bounding box coordinates of an apple. Similarly for DINOv you can provide a reference image and mask:

agent(
    "Can you detect all of the objects similar to the mask I've provided?",
    image="image.jpg",
    reference_data={"mask": "reference_mask.png", "image": "reference_image.png"},
)

Here, reference_mask.png and reference_image.png in reference_data could be any image with it's corresponding mask that is the object you want to detect in image.jpg. You can find a demo app to generate masks for DINOv here.

Tools

There are a variety of tools for the model or the user to use. Some are executed locally while others are hosted for you. You can also ask an LLM directly to build a tool for you. For example:

>>> import vision_agent as va
>>> llm = va.llm.OpenAILLM()
>>> detector = llm.generate_detector("Can you build a jar detector for me?")
>>> detector("jar.jpg")
[{"labels": ["jar",],
  "scores": [0.99],
  "bboxes": [
    [0.58, 0.2, 0.72, 0.45],
  ]
}]

Custom Tools

You can also add your own custom tools for your vision agent to use:

from vision_agent.tools import Tool, register_tool
@register_tool
class NumItems(Tool):
   name = "num_items_"
   description = "Returns the number of items in a list."
   usage = {
       "required_parameters": [{"name": "prompt", "type": "list"}],
       "examples": [
           {
               "scenario": "How many items are in this list? ['a', 'b', 'c']",
               "parameters": {"prompt": "['a', 'b', 'c']"},
           }
       ],
   }
   def __call__(self, prompt: list[str]) -> int:
       return len(prompt)

This will register it with the list of tools Vision Agent has access to. It will be able to pick it based on the tool description and use it based on the usage provided. You can find an example that creates a custom tool for template matching here.

Tool List

Tool	Description
CLIP	CLIP is a tool that can classify or tag any image given a set of input classes or tags.
ImageCaption	ImageCaption is a tool that can generate a caption for an image.
GroundingDINO	GroundingDINO is a tool that can detect arbitrary objects with inputs such as category names or referring expressions.
GroundingSAM	GroundingSAM is a tool that can detect and segment arbitrary objects with inputs such as category names or referring expressions.
DINOv	DINOv is a tool that can detect arbitrary objects with using a referring mask.
ExtractFrames	ExtractFrames extracts frames with motion from a video.
Crop	Crop crops an image given a bounding box and returns a file name of the cropped image.
BboxArea	BboxArea returns the area of the bounding box in pixels normalized to 2 decimal places.
SegArea	SegArea returns the area of the segmentation mask in pixels normalized to 2 decimal places.
BboxIoU	BboxIoU returns the intersection over union of two bounding boxes normalized to 2 decimal places.
SegIoU	SegIoU returns the intersection over union of two segmentation masks normalized to 2 decimal places.
BoxDistance	BoxDistance returns the minimum distance between two bounding boxes normalized to 2 decimal places.
MaskDistance	MaskDistance returns the minimum distance between two segmentation masks in pixel units
BboxContains	BboxContains returns the intersection of two boxes over the target box area. It is good for check if one box is contained within another box.
ExtractFrames	ExtractFrames extracts frames with motion from a video.
ZeroShotCounting	ZeroShotCounting returns the total number of objects belonging to a single class in a given image.
VisualPromptCounting	VisualPromptCounting returns the total number of objects belonging to a single class given an image and visual prompt.
VisualQuestionAnswering	VisualQuestionAnswering is a tool that can explain the contents of an image and answer questions about the image.
ImageQuestionAnswering	ImageQuestionAnswering is similar to VisualQuestionAnswering but does not rely on OpenAI and instead uses a dedicated model for the task.
OCR	OCR returns the text detected in an image along with the location.

It also has a basic set of calculate tools such as add, subtract, multiply and divide.

Azure Setup

If you want to use Azure OpenAI models, you can set the environment variable:

export AZURE_OPENAI_API_KEY="your-api-key"
export AZURE_OPENAI_ENDPOINT="your-endpoint"

You can then run Vision Agent using the Azure OpenAI models:

>>> import vision_agent as va
>>> agent = va.agent.VisionAgent(
>>>     task_model=va.llm.AzureOpenAILLM(),
>>>     answer_model=va.lmm.AzureOpenAILMM(),
>>>     reflection_model=va.lmm.AzureOpenAILMM(),
>>> )

antunsz/vision-agent