app.py

import gradio as gr
from transformers import pipeline
from transformers.image_utils import load_image

checkpoints = [
    'ustc-community/dfine_n_coco',
    'ustc-community/dfine_s_coco',
    'ustc-community/dfine_m_coco',
    'ustc-community/dfine_l_coco',
    'ustc-community/dfine_x_coco',
    'ustc-community/dfine_s_obj365',
    'ustc-community/dfine_m_obj365',
    'ustc-community/dfine_l_obj365',
    'ustc-community/dfine_x_obj365',
    'ustc-community/dfine_s_obj2coco',
    'ustc-community/dfine_m_obj2coco',
    'ustc-community/dfine_l_obj2coco_e25',
    'ustc-community/dfine_x_obj2coco',
]

def detect_objects(image, checkpoint, confidence_threshold=0.3, use_url=False, url=""):
    pipe = pipeline(
        "object-detection",
        model=checkpoint,
        image_processor=checkpoint,
        device="cpu",
    )

    if use_url and url:
        input_image = load_image(url)
    elif image is not None:
        input_image = image
    else:
        return None, gr.Markdown("**Error**: Please provide an image or URL.", visible=True)

    # Run detection
    results = pipe(input_image, threshold=confidence_threshold)

    # Get image dimensions for validation
    img_width, img_height = input_image.size

    # Prepare annotations in the format: list of (bounding_box, label)
    annotations = []
    for result in results:
        score = result["score"]
        if score < confidence_threshold:
            continue
        label = f"{result['label']} ({score:.2f})"
        box = result["box"]
        # Validate and convert box to (x1, y1, x2, y2)
        x1 = max(0, int(box["xmin"]))
        y1 = max(0, int(box["ymin"]))
        x2 = min(img_width, int(box["xmax"]))
        y2 = min(img_height, int(box["ymax"]))
        # Ensure valid box
        if x2 <= x1 or y2 <= y1:
            continue
        bounding_box = (x1, y1, x2, y2)
        annotations.append((bounding_box, label))

    # Handle empty annotations
    if not annotations:
        return (input_image, []), gr.Markdown(
            "**Warning**: No objects detected above the confidence threshold. Try lowering the threshold.",
            visible=True
        )

    # Return base image and annotations
    return (input_image, annotations), gr.Markdown(visible=False)

# Gradio interface
with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown(
        """
        # Real-Time Object Detection Demo
        Experience state-of-the-art object detection with USTC's Dfine models. Upload an image, provide a URL, or try an example below. Select a model and adjust the confidence threshold to see detections in real time!

        **Instructions**:
        - Upload an image or enter a URL.
        - Choose a model checkpoint from the dropdown.
        - Adjust the confidence threshold (0.1 to 1.0).
        - Click "Detect Objects" to view results, or select an example.
        - Use "Clear" to reset inputs and outputs.
        """,
        elem_classes="header-text"
    )
    
    with gr.Row():
        with gr.Column(scale=1, min_width=300):
            with gr.Group():
                image_input = gr.Image(
                    label="Upload Image",
                    type="pil",
                    sources=["upload", "webcam"],
                    interactive=True,
                    elem_classes="input-component",
                )
                use_url = gr.Checkbox(label="Use Image URL Instead", value=False)
                url_input = gr.Textbox(
                    label="Image URL",
                    placeholder="https://example.com/image.jpg",
                    visible=False,
                    elem_classes="input-component",
                )
                checkpoint = gr.Dropdown(
                    choices=checkpoints,
                    label="Select Model Checkpoint",
                    value=checkpoints[0],
                    elem_classes="input-component",
                )
                confidence_threshold = gr.Slider(
                    minimum=0.1,
                    maximum=1.0,
                    value=0.3,
                    step=0.1,
                    label="Confidence Threshold",
                    elem_classes="input-component",
                )
                with gr.Row():
                    detect_button = gr.Button(
                        "Detect Objects",
                        variant="primary",
                        elem_classes="action-button",
                    )
                    clear_button = gr.Button(
                        "Clear",
                        variant="secondary",
                        elem_classes="action-button",
                    )
        
        with gr.Column(scale=2):
            output_annotated = gr.AnnotatedImage(
                label="Detection Results",
                show_label=True,
                color_map=None,  # Let Gradio assign colors
                elem_classes="output-component",
            )
            error_message = gr.Markdown(visible=False, elem_classes="error-text")
    
    gr.Examples(
        examples=[
            ["./image.jpg", False, "", checkpoints[0], 0.3],
            [None, True, "https://live.staticflickr.com/65535/33021460783_1646d43c54_b.jpg", checkpoints[0], 0.3],
        ],
        inputs=[image_input, use_url, url_input, checkpoint, confidence_threshold],
        outputs=[output_annotated, error_message],
        fn=detect_objects,
        cache_examples=False,  # Avoid caching due to model size
        label="Select an example to run the model",
    )

    # Dynamic visibility for URL input
    use_url.change(
        fn=lambda x: gr.update(visible=x),
        inputs=use_url,
        outputs=url_input,
    )

    # Clear button functionality
    clear_button.click(
        fn=lambda: (
            None,  # image_input
            False,  # use_url
            "",  # url_input
            checkpoints[0],  # checkpoint
            0.3,  # confidence_threshold
            None,  # output_annotated
            gr.Markdown(visible=False),  # error_message
        ),
        outputs=[
            image_input,
            use_url,
            url_input,
            checkpoint,
            confidence_threshold,
            output_annotated,
            error_message,
        ],
    )

    # Detect button event
    detect_button.click(
        fn=detect_objects,
        inputs=[image_input, checkpoint, confidence_threshold, use_url, url_input],
        outputs=[output_annotated, error_message],
    )

if __name__ == "__main__":
    demo.launch()