app.py

import gradio as gr
import torch
from PIL import Image, ImageDraw, ImageFont
from transformers import DetrImageProcessor, DetrForObjectDetection
from pathlib import Path


# Load DETR model and processor from Hugging Face
model_name = "facebook/detr-resnet-50"
processor = DetrImageProcessor.from_pretrained(model_name)
model = DetrForObjectDetection.from_pretrained(model_name)

# Load font
font_path = Path("assets/fonts/arial.ttf")
if not font_path.exists():
    # If the font file does not exist, use the default PIL font
    print(f"Font file {font_path} not found. Using default font.")
    font = ImageFont.load_default()
else:
    font = ImageFont.truetype(str(font_path), size=100)

print(f"CUDA is available: {torch.cuda.is_available()}")

# Main function: takes an image and returns it with boxes and labels
def detect_objects(image):
    inputs = processor(images=image, return_tensors="pt")
    outputs = model(**inputs)

    # Convert model output to usable detection results
    target_sizes = torch.tensor([image.size[::-1]])
    results = processor.post_process_object_detection(
        outputs, threshold=0.9, target_sizes=target_sizes
    )[0]

    # Draw bounding boxes and labels on a copy of the image
    image_with_boxes = image.copy()
    draw = ImageDraw.Draw(image_with_boxes)

    for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
        box = [round(x, 2) for x in box.tolist()]
        draw.rectangle(box, outline="red", width=3)

        # Prepare label text
        label_text = f"{model.config.id2label[label.item()]}: {round(score.item(), 2)}"

        # Measure text size
        text_bbox = draw.textbbox((0, 0), label_text, font=font)
        text_width = text_bbox[2] - text_bbox[0]
        text_height = text_bbox[3] - text_bbox[1]

        # Set background rectangle for text
        text_background = [
            box[0], box[1] - text_height,
                    box[0] + text_width, box[1]
        ]
        draw.rectangle(text_background, fill="black")  # Background
        draw.text((box[0], box[1] - text_height), label_text, fill="white", font=font)

    return image_with_boxes


with gr.Blocks() as app:
    with gr.Row():
        gr.Markdown("## Object Detection App\nUpload an image to detect objects using Facebook's DETR model.")
    with gr.Row():
        input_image = gr.Image(type="pil", label="Input Image")
        output_image = gr.Image(label="Detected Objects")
    with gr.Row():
        button = gr.Button("Detect Objects")

    button.click(fn=detect_objects, inputs=input_image, outputs=output_image)

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