app.py

import gradio as gr
from ultralytics import YOLOv10
import cv2
import torch
import os
import spaces


model = YOLOv10.from_pretrained('jameslahm/yolov10x')

# Define activity categories based on detected objects
activity_categories = {
    "Working": ["laptop", "computer", "keyboard", "office chair"],
    "Meal Time": ["fork", "spoon", "plate", "food"],
    "Exercise": ["dumbbell", "bicycle", "yoga mat", "treadmill"],
    "Outdoors": ["car", "tree", "bicycle", "road"],
    # Add more categories and objects as needed
}

# Function to map detected objects to categorized activities
def categorize_activity(detected_objects):
    categorized_activities = {}
    
    for activity, objects in activity_categories.items():
        if any(obj in detected_objects for obj in objects):
            if activity not in categorized_activities:
                categorized_activities[activity] = []
            categorized_activities[activity].append(detected_objects)
    
    return categorized_activities

# Function to process the video, detect objects, and generate a categorized journal with images
@spaces.GPU
def generate_journal_with_images(video_path):
    model.to("cuda")
    cap = cv2.VideoCapture(video_path)
    journal_entries = {}
    saved_images = []
    frame_count = 0
    output_folder = "detected_frames"
    os.makedirs(output_folder, exist_ok=True)  # Create folder to store images
    
    while cap.isOpened():
        ret, frame = cap.read()
        if not ret:
            break
        
        frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
        
        # Make predictions using YOLOv10 on the current frame
        results = model.predict(source=frame_rgb, device=device)
        
        # Draw bounding boxes on the frame
        results.render()  # Render the results on the image (this modifies the frame in-place)
        
        # Save the image with bounding boxes
        frame_filename = os.path.join(output_folder, f"frame_{frame_count}.jpg")
        cv2.imwrite(frame_filename, frame_rgb[:, :, ::-1])  # Convert back to BGR for saving
        saved_images.append(frame_filename)
        
        # Extract labels (class indices) and map them to class names
        detected_objects = [model.names[int(box.cls)] for box in results.boxes]
        
        # Get current timestamp in the video
        timestamp = cap.get(cv2.CAP_PROP_POS_MSEC) / 1000  # Convert ms to seconds
        
        # Categorize the detected objects into activities
        activity_summary = categorize_activity(detected_objects)
        
        # Store the activities with their timestamp
        for activity, objects in activity_summary.items():
            if activity not in journal_entries:
                journal_entries[activity] = []
            journal_entries[activity].append((f"At {timestamp:.2f} seconds: {', '.join(objects[0])}", frame_filename))
        
        frame_count += 1
    
    cap.release()
    
    # Create a formatted journal output
    formatted_journal = []
    for activity, entries in journal_entries.items():
        formatted_journal.append(f"**{activity}:**")
        for entry, image_path in entries:
            formatted_journal.append((entry, image_path))
    
    return formatted_journal

# Gradio interface for uploading video and generating journal with images
def display_journal_with_images(video):
    journal_with_images = generate_journal_with_images(video)
    
    # Create the final display with text and images
    display_items = []
    for entry, image_path in journal_with_images:
        display_items.append((entry, image_path))
    
    return display_items

# Define Gradio Blocks for custom display
with gr.Blocks() as iface:
    video_input = gr.Video(label="Upload Video")
    output_gallery = gr.Gallery(label="Generated Daily Journal with Images").style(grid=[2], height='auto')
    run_button = gr.Button("Generate Journal")
    
    run_button.click(fn=display_journal_with_images, inputs=video_input, outputs=output_gallery)

iface.launch()