feat: ✨ Add xyxy_xywh function and from_transformers method for KeyPoints class added

[onuralpszr]

🚀 from_transformers method for KeyPoints class added to supervision 🎯

✨ New Functionality

• 🆕 Bounding Box Conversion

  • Added xyxy_xywh function to convert bounding box coordinates from (x_min, y_min, x_max, y_max) format to (x, y, width, height) format.
  • 📍 Location: supervision/detection/utils.py

• 🆕 Transformers Model Integration

  • Introduced from_transformers class method in supervision/keypoint/core.py.
  • Enables creating KeyPoints objects from Transformers inference results.
  • 📝 Includes detailed documentation and usage examples.

🔧 Updates to __init__.py

• ✅ Updated supervision/__init__.py to include the xyxy_to_xywh function in imports.
• ✅ Added xyxy_to_xywh to the __all__ list to ensure accessibility when the module is imported.

💡 Feel free to test it using the provided Colab notebook!

[SkalskiP]

Hi @onuralpszr 👋🏻 thanks a lot for working on this feature. Could you please:

• Add unit tests for xyxy_xywh
• Add xyxy_xywh to docs/detection/utils.md
• Slightly update naming convention in from_transformers docstring example

from PIL import Image
import requests
import supervision as sv
import torch
from transformers import (
    AutoProcessor,
    RTDetrForObjectDetection,
    VitPoseForPoseEstimation,
)

device = "cuda" if torch.cuda.is_available() else "cpu"
image = Image.open(<SOURCE_IMAGE_PATH>)

DETECTION_MODEL_ID = "PekingU/rtdetr_r50vd_coco_o365"

detection_processor = AutoProcessor.from_pretrained(DETECTION_MODEL_ID, use_fast=True)
detection_model = RTDetrForObjectDetection.from_pretrained(DETECTION_MODEL_ID, device_map=DEVICE)

inputs = detection_processor(images=frame, return_tensors="pt").to(DEVICE)

with torch.no_grad():
    outputs = detection_model(**inputs)

target_size = torch.tensor([(frame.height, frame.width)])
results = detection_processor.post_process_object_detection(
    outputs, target_sizes=target_size, threshold=0.3)

detections = sv.Detections.from_transformers(results[0])
boxes = sv.xyxy_xywh(detections[detections.class_id == 0].xyxy)

POSE_ESTIMATION_MODEL_ID = "usyd-community/vitpose-base-simple"

pose_estimation_processor = AutoProcessor.from_pretrained(POSE_ESTIMATION_MODEL_ID)
pose_estimation_model = VitPoseForPoseEstimation.from_pretrained(
    POSE_ESTIMATION_MODEL_ID, device_map=DEVICE)

inputs = pose_estimation_processor(frame, boxes=[boxes], return_tensors="pt").to(DEVICE)

with torch.no_grad():
    outputs = pose_estimation_model(**inputs)

results = pose_estimation_processor.post_process_pose_estimation(outputs, boxes=[boxes])
key_point = sv.KeyPoints.from_transformers(results[0])

[SkalskiP]

Awesome @onuralpszr ! Merging! 🔥

[onuralpszr]

Awesome @onuralpszr ! Merging! 🔥

Thank you very much ! :)