app.py

import gradio as gr
import spaces
import supervision as sv
from rfdetr import RFDETRBase, RFDETRLarge
from rfdetr.util.coco_classes import COCO_CLASSES

from utils.video import create_directory

MARKDOWN = """
# RF-DETR 🔥

<div style="display: flex; align-items: center; gap: 8px;">
  <a href="https://colab.research.google.com/github/roboflow-ai/notebooks/blob/main/notebooks/how-to-finetune-rf-detr-on-detection-dataset.ipynb">
    <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="colab" />
  </a>
  <a href="https://blog.roboflow.com/rf-detr">
    <img src="https://raw.githubusercontent.com/roboflow-ai/notebooks/main/assets/badges/roboflow-blogpost.svg" alt="roboflow" />
  </a>
  <a href="https://github.com/roboflow/rf-detr">
    <img src="https://badges.aleen42.com/src/github.svg" alt="roboflow" />
  </a>
</div>

RF-DETR is a real-time, transformer-based object detection model architecture developed 
by [Roboflow](https://roboflow.com/) and released under the Apache 2.0 license.
"""

IMAGE_EXAMPLES = [
    ['https://media.roboflow.com/supervision/image-examples/people-walking.png', 0.3, 728, "large"],
    ['https://media.roboflow.com/supervision/image-examples/vehicles.png', 0.3, 728, "large"],
    ['https://media.roboflow.com/notebooks/examples/dog-2.jpeg', 0.5, 560, "base"],
]

COLOR = sv.ColorPalette.from_hex([
    "#ffff00", "#ff9b00", "#ff8080", "#ff66b2", "#ff66ff", "#b266ff",
    "#9999ff", "#3399ff", "#66ffff", "#33ff99", "#66ff66", "#99ff00"
])

VIDEO_SCALE_FACTOR = 0.5
VIDEO_TARGET_DIRECTORY = "tmp"
create_directory(directory_path=VIDEO_TARGET_DIRECTORY)


@spaces.GPU()
def inference(image, confidence: float, resolution: int, checkpoint: str):
    model_class = RFDETRBase if checkpoint == "base" else RFDETRLarge
    model = model_class(resolution=resolution)
    detections = model.predict(image, threshold=confidence)

    text_scale = sv.calculate_optimal_text_scale(resolution_wh=image.size)
    thickness = sv.calculate_optimal_line_thickness(resolution_wh=image.size)

    bbox_annotator = sv.BoxAnnotator(color=COLOR, thickness=thickness)
    label_annotator = sv.LabelAnnotator(
        color=COLOR,
        text_color=sv.Color.BLACK,
        text_scale=text_scale,
        smart_position=True
    )

    labels = [
        f"{COCO_CLASSES[class_id]} {confidence:.2f}"
        for class_id, confidence
        in zip(detections.class_id, detections.confidence)
    ]

    annotated_image = image.copy()
    annotated_image = bbox_annotator.annotate(annotated_image, detections)
    annotated_image = label_annotator.annotate(annotated_image, detections, labels)
    return annotated_image

with gr.Blocks() as demo:
    gr.Markdown(MARKDOWN)
    with gr.Row():
        with gr.Column():
            input_image = gr.Image(
                label="Input Image",
                image_mode='RGB',
                type='pil',
                height=600
            )
            confidence_slider = gr.Slider(
                label="Confidence",
                minimum=0.0,
                maximum=1.0,
                step=0.05,
                value=0.5,
            )
            resolution_slider = gr.Slider(
                label="Inference resolution",
                minimum=560,
                maximum=1120,
                step=56,
                value=728,
            )
            with gr.Row():
                checkpoint_dropdown = gr.Dropdown(
                    label="Checkpoint",
                    choices=["base", "large"],
                    value="base"
                )
                submit_button = gr.Button("Submit")
        with gr.Column():
            output_image = gr.Image(
                label="Input Image",
                image_mode='RGB',
                type='pil',
                height=600
            )
    gr.Examples(
        fn=inference,
        examples=IMAGE_EXAMPLES,
        inputs=[input_image, confidence_slider, resolution_slider, checkpoint_dropdown],
        outputs=output_image
    )

    submit_button.click(
        inference,
        inputs=[input_image, confidence_slider, resolution_slider, checkpoint_dropdown],
        outputs=output_image
    )

demo.launch(debug=False, show_error=True)