app.py

import gradio as gr
import cv2
import numpy as np
from ultralytics import YOLO

# Load YOLO model
model = YOLO('yolo11s-earth.pt')  # Load your model

# Default classes
default_classes = [
    'airplane', 'airport', 'baseballfield', 'basketballcourt', 'bridge',
    'chimney', 'dam', 'Expressway-Service-area', 'Expressway-toll-station',
    'golffield', 'groundtrackfield', 'harbor', 'overpass', 'ship',
    'stadium', 'storagetank', 'tenniscourt', 'trainstation', 'vehicle',
    'windmill'
]

def process_frame(frame, classes_input):
    # Process user input classes
    if classes_input and classes_input.strip():
        classes_list = [cls.strip() for cls in classes_input.split(',')]
        # Validate classes_list
        for cls in classes_list:
            if not isinstance(cls, str):
                print("Invalid class name:", cls)
                continue
        model.set_classes(classes_list)  # Set model classes
    else:
        # Use default classes if no input or input is empty
        model.set_classes(default_classes)
    
    # Copy frame to a writable array
    frame = frame.copy()
    
    # Resize image to speed up processing (optional)
    h, w = frame.shape[:2]
    new_size = (640, int(h * (640 / w))) if w > h else (int(w * (640 / h)), 640)
    resized_frame = cv2.resize(frame, new_size)
    
    # Convert image format
    rgb_frame = cv2.cvtColor(resized_frame, cv2.COLOR_BGR2RGB)
    
    # Use model for detection
    results = model.predict(rgb_frame)
    
    # Draw detection results
    for result in results:
        boxes = result.boxes
        for box in boxes:
            x1, y1, x2, y2 = box.xyxy[0]
            conf = box.conf[0]
            cls = box.cls[0]
            try:
                class_name = model.names[int(cls)]
            except (IndexError, TypeError) as e:
                print(f"Error accessing model.names: {e}")
                class_name = "Unknown"  # Provide a default value
            
            # Adjust coordinates to original image size
            x1 = int(x1 * w / new_size[0])
            y1 = int(y1 * h / new_size[1])
            x2 = int(x2 * w / new_size[0])
            y2 = int(y2 * h / new_size[1])
            
            # Draw bounding box and label
            cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), 2)
            cv2.putText(frame, f'{class_name}:{conf:.2f}', (x1, y1 - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.9, (36, 255, 12), 2)
    
    return frame

def main():
    # Create Gradio interface
    with gr.Blocks() as demo:
        gr.Markdown("# YOLO11s-Earth open vocabulary detection (DIOR finetuning)")
        with gr.Row():
            cam_input = gr.Image(type="numpy", sources=["webcam"], streaming=True, label="Webcam")
            classes_input = gr.Textbox(label="New classes (comma-separated)", placeholder="e.g.: airplane, airport, tennis court")
        output = gr.Image(label="Results", type="numpy", height=480)  # Set height to 480
        
        cam_input.stream(
            process_frame,
            inputs=[cam_input, classes_input],
            outputs=output
        )
    
    # Launch Gradio app
    demo.launch()

if __name__ == "__main__":
    main()