app.py

import os
import gradio as gr
from transformers import pipeline, DetrForObjectDetection, DetrConfig, DetrImageProcessor
import numpy as np
import cv2
from PIL import Image

# Pre-load the base configuration and models (without setting a threshold yet)
base_config = DetrConfig.from_pretrained("facebook/detr-resnet-50")
base_model = DetrForObjectDetection.from_pretrained("facebook/detr-resnet-50", config=base_config)
base_processor = DetrImageProcessor.from_pretrained("facebook/detr-resnet-50")

def load_model(threshold):
    # Adjust the configuration for the current threshold
    config = DetrConfig.from_pretrained("facebook/detr-resnet-50", threshold=threshold)
    # Create a new model instance with the updated configuration
    model = DetrForObjectDetection.from_pretrained("facebook/detr-resnet-50", config=config)
    # Image processor does not need to be re-loaded
    return pipeline(task='object-detection', model=model, image_processor=base_processor)

# Initialize the pipeline with a default threshold
od_pipe = load_model(0.25)  # Set a default threshold here

def draw_detections(image, detections):
    np_image = np.array(image)
    np_image = cv2.cvtColor(np_image, cv2.COLOR_RGB2BGR)
    
    for detection in detections:
        score = detection['score']
        label = detection['label']
        box = detection['box']
        x_min, y_min = box['xmin'], box['ymin']
        x_max, y_max = box['xmax'], box['ymax']
        # Draw rectangles and text with a larger font
        cv2.rectangle(np_image, (x_min, y_min), (x_max, y_max), (0, 255, 0), 2)
        label_text = f'{label} {score:.2f}'
        # Increase the font size and text thickness
        cv2.putText(np_image, label_text, (x_min, y_min - 10), cv2.FONT_HERSHEY_SIMPLEX, 1.5, (255, 255, 255), 4)
    
    # Convert BGR to RGB for displaying
    final_image = cv2.cvtColor(np_image, cv2.COLOR_BGR2RGB)
    final_pil_image = Image.fromarray(final_image)
    return final_pil_image

def get_pipeline_prediction(threshold, pil_image):
    global od_pipe
    od_pipe = load_model(threshold)  # reload model with the specified threshold
    try:
        if not isinstance(pil_image, Image.Image):
            pil_image = Image.fromarray(np.array(pil_image).astype('uint8'), 'RGB')
        result = od_pipe(pil_image)
        processed_image = draw_detections(pil_image, result)
        return processed_image, result
    except Exception as e:
        return pil_image, {"error": str(e)}

with gr.Blocks() as demo:
    with gr.Row():
        with gr.Column():
            gr.Markdown("## Object Detection")
            inp_image = gr.Image(label="Upload your image here")
            threshold_slider = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, value=0.25, label="Detection Threshold")
            run_button = gr.Button("Detect Objects")
        with gr.Column():
            with gr.Tab("Annotated Image"):
                output_image = gr.Image()
            with gr.Tab("Detection Results"):
                output_data = gr.JSON()

    run_button.click(get_pipeline_prediction, inputs=[threshold_slider, inp_image], outputs=[output_image, output_data])

demo.launch()