​
Overview

OwlV2 (Open-World Language-guided Visual representation) is a model that combines natural language processing with computer vision to enable open-world understanding and object detection based on textual descriptions.

OwlV2()

The OwlV2 class is a wrapper around the pre-trained OWLv2 model from Hugging Face. It provides an interface for loading the model, detecting objects in an image based on a list of candidate classes, and returning the detections as bounding boxes.

​
Methods

OwlV2.detect()
  • detect(self, image: Image.Image, classes: List[str]) -> Detections: Detects objects in the given image based on the provided list of classes. Returns the detections as a Detections object containing bounding box coordinates, confidence scores, and class IDs.

​
When to use?

OwlV2 excels in language understanding for niche bounding box tasks. It was pretrained on a web-scale corpus of 10B+ images, making it particularly effective for detecting specific or unusual objects that are not typically found in standard datasets like COCO.

OwlV2 does often struggle with the problem of overpredicting bounding boxes or predicting multiple boxes for a single instance, so its reccomended that you chain this model with Non Max Suppression to get cleaner results.

​
Example Usage

wget -O kites.jpg "https://github.com/overeasy-sh/overeasy/blob/main/examples/kites.jpg?raw=true"

from overeasy import *
from PIL import Image
workflow = Workflow([
  BoundingBoxSelectAgent(classes=["butterfly kite"], model=OwlV2()),
  NMSAgent(iou_threshold=0.8, score_threshold=0.6),
])

image = Image.open("./kites.jpg")
results, graph = workflow.execute(image)
workflow.visualize(graph)

For more information on OwlV2, refer to the original paper.

Grounding DINOYOLO-World