app.py

import gradio as gr
import PIL.Image
import torch
from transformers import AutoModelForZeroShotObjectDetection, AutoProcessor
from transformers.image_utils import load_image

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"


class Detector:
    def __init__(self, model_id: str):
        self.device = DEVICE
        self.processor = AutoProcessor.from_pretrained(model_id)
        self.model = AutoModelForZeroShotObjectDetection.from_pretrained(model_id).to(
            self.device
        )

    def detect(
        self,
        image: PIL.Image.Image,
        text_labels: list[str],
        threshold: float = 0.4,
    ):
        inputs = self.processor(
            images=image, text=[text_labels], return_tensors="pt"
        ).to(self.device)

        with torch.no_grad():
            outputs = self.model(**inputs)

        results = self.processor.post_process_grounded_object_detection(
            outputs, threshold=threshold, target_sizes=[(image.height, image.width)]
        )

        detections = []
        result = results[0]
        for box, score, labels in zip(
            result["boxes"], result["scores"], result["labels"]
        ):
            box = [round(x, 2) for x in box.tolist()]
            detections.append(
                dict(
                    label=labels,
                    confidence=round(score.item(), 3),
                    box=box,
                )
            )
        return detections


models = dict(
    tiny=Detector("iSEE-Laboratory/llmdet_tiny"),
    base=Detector("iSEE-Laboratory/llmdet_base"),
    large=Detector("iSEE-Laboratory/llmdet_large"),
)


def _postprocess(detections):
    annotations = []
    for detection in detections:
        box = detection["box"]
        mask = (int(box[0]), int(box[1]), int(box[2]), int(box[3]))
        label = f"{detection['label']} ({detection['confidence']:.2f})"
        annotations.append((mask, label))
    return annotations


def detect_objects(image, labels, confidence_threshold):
    labels = [label.strip() for label in labels.split(",")]

    detections = []
    for model_name in models.keys():
        detection = models[model_name].detect(
            image,
            labels,
            threshold=confidence_threshold,
        )
        detections.append(_postprocess(detection))

    return tuple((image, det) for det in detections)


with gr.Blocks() as demo:
    gr.Markdown("# [LLMDet](https://arxiv.org/abs/2501.18954) Arena ✨")

    with gr.Row():
        with gr.Column():
            gr.Markdown("## Input Image")

            image_url = "http://images.cocodataset.org/val2017/000000039769.jpg"
            image = load_image(image_url)
            image_input = gr.Image(type="pil", image_mode="RGB", value=image)

        with gr.Column():
            gr.Markdown("## Settings")

            confidence_slider = gr.Slider(
                0,
                1,
                value=0.4,
                step=0.01,
                interactive=True,
                label="Confidence threshold:",
            )

            labels = ["a cat", "a remote control"]

            text_input = gr.Textbox(
                label="Object labels (comma separated):",
                placeholder=",".join(labels),
                lines=1,
                value=",".join(labels),
            )

    with gr.Row():
        detect_button = gr.Button("Run Object Detection")

    with gr.Row():
        gr.Markdown("## Output Annotated Images")

    with gr.Row():
        output_annotated_image_tiny = gr.AnnotatedImage(label="TINY")
        output_annotated_image_base = gr.AnnotatedImage(label="BASE")
        output_annotated_image_large = gr.AnnotatedImage(label="LARGE")

    # Connect the button to the detection function
    detect_button.click(
        fn=detect_objects,
        inputs=[image_input, text_input, confidence_slider],
        outputs=[
            output_annotated_image_tiny,
            output_annotated_image_base,
            output_annotated_image_large,
        ],
    )

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