app.py

import spaces
import gradio as gr
from PIL import Image, ImageDraw, ImageFont
from ultralytics import YOLO
from huggingface_hub import hf_hub_download
import cv2
import tempfile
import numpy as np

def download_model(model_filename):
    """
    Downloads a YOLO model from the Hugging Face Hub.

    This function fetches a specified YOLO model file from the
    'atalaydenknalbant/Yolov13' repository on the Hugging Face Hub.

    Args:
        model_filename (str): The name of the model file to download
                              (e.g., 'yolov13n.pt').

    Returns:
        str: The local path to the downloaded model file.
    """
    return hf_hub_download(repo_id="atalaydenknalbant/Yolov13", filename=model_filename)

@spaces.GPU
def yolo_inference(input_type, image, video, model_id, conf_threshold, iou_threshold, max_detection):
    """
    Performs object detection inference using a YOLOv13 model on either an image or a video.

    This function downloads the specified YOLO model, then applies it to the
    provided input. For images, it returns an annotated image. For videos, it
    processes each frame and returns an annotated video. Error handling for
    missing inputs is included, returning blank outputs with messages.

    Args:
        input_type (str): Specifies the input type, either "Image" or "Video".
        image (PIL.Image.Image or None): The input image if `input_type` is "Image".
                                         None otherwise.
        video (str or None): The path to the input video file if `input_type` is "Video".
                             None otherwise.
        model_id (str): The identifier of the YOLO model to use (e.g., 'yolov13n.pt').
        conf_threshold (float): The confidence threshold for object detection.
                                Detections with lower confidence are discarded.
        iou_threshold (float): The Intersection over Union (IoU) threshold for
                               Non-Maximum Suppression (NMS).
        max_detection (int): The maximum number of detections to return per image or frame.

    Returns:
        tuple: A tuple containing two elements:
            - PIL.Image.Image or None: The annotated image if `input_type` was "Image",
                                       otherwise None.
            - str or None: The path to the annotated video file if `input_type` was "Video",
                           otherwise None.
    """
    model_path = download_model(model_id)

    if input_type == "Image":
        if image is None:
            width, height = 640, 480
            blank_image = Image.new("RGB", (width, height), color="white")
            draw = ImageDraw.Draw(blank_image)
            message = "No image provided"
            font = ImageFont.load_default(size=40)
            bbox = draw.textbbox((0, 0), message, font=font)
            text_width = bbox[2] - bbox[0]
            text_height = bbox[3] - bbox[1]
            text_x = (width - text_width) / 2
            text_y = (height - text_height) / 2
            draw.text((text_x, text_y), message, fill="black", font=font)
            return blank_image, None

        model = YOLO(model_path)
        results = model.predict(
            source=image,
            conf=conf_threshold,
            iou=iou_threshold,
            imgsz=640,
            max_det=max_detection,
            show_labels=True,
            show_conf=True,
        )
        for r in results:
            image_array = r.plot()
            annotated_image = Image.fromarray(image_array[..., ::-1])
        return annotated_image, None

    elif input_type == "Video":
        if video is None:
            width, height = 640, 480
            blank_image = Image.new("RGB", (width, height), color="white")
            draw = ImageDraw.Draw(blank_image)
            message = "No video provided"
            font = ImageFont.load_default(size=40)
            bbox = draw.textbbox((0, 0), message, font=font)
            text_width = bbox[2] - bbox[0]
            text_height = bbox[3] - bbox[1]
            text_x = (width - text_width) / 2
            text_y = (height - text_height) / 2
            draw.text((text_x, text_y), message, fill="black", font=font)
            temp_video_file = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False).name
            fourcc = cv2.VideoWriter_fourcc(*"mp4v")
            out = cv2.VideoWriter(temp_video_file, fourcc, 1, (width, height))
            frame = cv2.cvtColor(np.array(blank_image), cv2.COLOR_RGB2BGR)
            out.write(frame)
            out.release()
            return None, temp_video_file

        model = YOLO(model_path)
        cap = cv2.VideoCapture(video)
        fps = cap.get(cv2.CAP_PROP_FPS) if cap.get(cv2.CAP_PROP_FPS) > 0 else 25
        frames = []
        while True:
            ret, frame = cap.read()
            if not ret:
                break
            pil_frame = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
            results = model.predict(
                source=pil_frame,
                conf=conf_threshold,
                iou=iou_threshold,
                imgsz=640,
                max_det=max_detection,
                show_labels=True,
                show_conf=True,
            )
            for r in results:
                annotated_frame_array = r.plot()
                annotated_frame = cv2.cvtColor(annotated_frame_array, cv2.COLOR_BGR2RGB)
            frames.append(annotated_frame)
        cap.release()
        if not frames:
            return None, None

        height_out, width_out, _ = frames[0].shape
        temp_video_file = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False).name
        fourcc = cv2.VideoWriter_fourcc(*"mp4v")
        out = cv2.VideoWriter(temp_video_file, fourcc, fps, (width_out, height_out))
        for f in frames:
            f_bgr = cv2.cvtColor(f, cv2.COLOR_RGB2BGR)
            out.write(f_bgr)
        out.release()
        return None, temp_video_file

    return None, None

def update_visibility(input_type):
    """
    Adjusts the visibility of Gradio components based on the selected input type.

    This function dynamically shows or hides the image and video input/output
    components in the Gradio interface to ensure only relevant fields are visible.

    Args:
        input_type (str): The selected input type, either "Image" or "Video".

    Returns:
        tuple: A tuple of `gr.update` objects for the visibility of:
               (image input, video input, image output, video output).
    """
    if input_type == "Image":
        return gr.update(visible=True), gr.update(visible=False), gr.update(visible=True), gr.update(visible=False)
    else:
        return gr.update(visible=False), gr.update(visible=True), gr.update(visible=False), gr.update(visible=True)

def yolo_inference_for_examples(image, model_id, conf_threshold, iou_threshold, max_detection):
    """
    Wrapper function for `yolo_inference` specifically for Gradio examples that use images.

    This function simplifies the `yolo_inference` call for the `gr.Examples` component,
    ensuring only image-based inference is performed for predefined examples.

    Args:
        image (PIL.Image.Image): The input image for the example.
        model_id (str): The identifier of the YOLO model to use.
        conf_threshold (float): The confidence threshold.
        iou_threshold (float): The IoU threshold.
        max_detection (int): The maximum number of detections.

    Returns:
        PIL.Image.Image or None: The annotated image. Returns None if no image is processed.
    """
    annotated_image, _ = yolo_inference(
        input_type="Image",
        image=image,
        video=None,
        model_id=model_id,
        conf_threshold=conf_threshold,
        iou_threshold=iou_threshold,
        max_detection=max_detection
    )
    return annotated_image

with gr.Blocks() as app:
    gr.Markdown("# Yolo13: Object Detection")
    gr.Markdown("Upload an image or video for inference using the latest YOLOv13 models.")
    gr.Markdown("📝 **Note:** Better-trained models will be deployed as they become available.")
    with gr.Accordion("Paper and Citation", open=False):
        gr.Markdown("""
        This application is based on the research from the paper: **YOLOv13: Real-Time Object Detection with Hypergraph-Enhanced Adaptive Visual Perception**.

        - **Authors:** Mengqi Lei, Siqi Li, Yihong Wu, et al.
        - **Preprint Link:** [https://arxiv.org/abs/2506.17733](https://arxiv.org/abs/2506.17733)

        **BibTeX:**
        ```
        @article{yolov13,
          title={YOLOv13: Real-Time Object Detection with Hypergraph-Enhanced Adaptive Visual Perception},
          author={Lei, Mengqi and Li, Siqi and Wu, Yihong and et al.},
          journal={arXiv preprint arXiv:2506.17733},
          year={2025}
        }
        ```
        """)

    with gr.Row():
        with gr.Column():
            image = gr.Image(type="pil", label="Image", visible=True)
            video = gr.Video(label="Video", visible=False)
            input_type = gr.Radio(
                choices=["Image", "Video"],
                value="Image",
                label="Input Type",
            )
            model_id = gr.Dropdown(
                label="Model Name",
                choices=[
                    'yolov13n.pt', 'yolov13s.pt', 'yolov13l.pt', 'yolov13x.pt',
                ],
                value="yolov13n.pt",
            )
            conf_threshold = gr.Slider(minimum=0, maximum=1, value=0.25, label="Confidence Threshold")
            iou_threshold = gr.Slider(minimum=0, maximum=1, value=0.45, label="IoU Threshold")
            max_detection = gr.Slider(minimum=1, maximum=300, step=1, value=300, label="Max Detection")
            infer_button = gr.Button("Detect Objects")
        with gr.Column():
            output_image = gr.Image(type="pil", show_label=False, show_share_button=False, visible=True)
            output_video = gr.Video(show_label=False, show_share_button=False, visible=False)
            gr.DeepLinkButton()

    input_type.change(
        fn=update_visibility,
        inputs=input_type,
        outputs=[image, video, output_image, output_video],
    )

    infer_button.click(
        fn=yolo_inference,
        inputs=[input_type, image, video, model_id, conf_threshold, iou_threshold, max_detection],
        outputs=[output_image, output_video],
    )

    gr.Examples(
        examples=[
            ["zidane.jpg", "yolov13s.pt", 0.35, 0.45, 300],
            ["bus.jpg", "yolov13l.pt", 0.35, 0.45, 300],
            ["yolo_vision.jpg", "yolov13x.pt", 0.35, 0.45, 300],
        ],
        fn=yolo_inference_for_examples,
        inputs=[image, model_id, conf_threshold, iou_threshold, max_detection],
        outputs=[output_image],
        label="Examples (Images)",
    )

if __name__ == '__main__':
    app.launch(mcp_server=True)