app.py

import streamlit as st
import json
import faiss
import numpy as np
from sentence_transformers import SentenceTransformer
import base64
from PIL import Image
import io
import cv2
from insightface.app import FaceAnalysis
from moviepy.editor import VideoFileClip
from sklearn.cluster import DBSCAN
from sklearn.decomposition import PCA
import plotly.graph_objs as go
from collections import defaultdict



# Load models
@st.cache_resource
def load_models():
    text_model = SentenceTransformer("all-MiniLM-L6-v2")
    image_model = SentenceTransformer("clip-ViT-B-32")
    face_app = FaceAnalysis(providers=['CPUExecutionProvider'])
    face_app.prepare(ctx_id=0, det_size=(640, 640))
    return text_model, image_model, face_app

text_model, image_model, face_app = load_models()

# Load data
@st.cache_data
def load_data(video_id, output_dir):
    with open(f"{output_dir}/{video_id}_summary.json", "r") as f:
        summary = json.load(f)
    with open(f"{output_dir}/{video_id}_transcription.json", "r") as f:
        transcription = json.load(f)
    with open(f"{output_dir}/{video_id}_text_metadata.json", "r") as f:
        text_metadata = json.load(f)
    with open(f"{output_dir}/{video_id}_image_metadata.json", "r") as f:
        image_metadata = json.load(f)
    with open(f"{output_dir}/{video_id}_face_metadata.json", "r") as f:
        face_metadata = json.load(f)
    face_index = faiss.read_index(f"{output_dir}/{video_id}_face_index.faiss")
    return summary, transcription, text_metadata, image_metadata, face_metadata, face_index

video_id = "IMFUOexuEXw"
output_dir = "./"
video_path = "avengers_interview.mp4"
summary, transcription, text_metadata, image_metadata, face_metadata, face_index = load_data(video_id, output_dir)

# Load FAISS indexes
@st.cache_resource
def load_indexes(video_id, output_dir):
    text_index = faiss.read_index(f"{output_dir}/{video_id}_text_index.faiss")
    image_index = faiss.read_index(f"{output_dir}/{video_id}_image_index.faiss")
    return text_index, image_index

text_index, image_index = load_indexes(video_id, output_dir)
def create_comprehensive_face_summary(face_index, face_metadata, eps=0.5, min_samples=3, top_k=5):
    # Extract face embeddings
    face_embeddings = face_index.reconstruct_n(0, face_index.ntotal)
    
    # Normalize embeddings
    face_embeddings = face_embeddings / np.linalg.norm(face_embeddings, axis=1)[:, np.newaxis]
    
    # Perform DBSCAN clustering
    clustering = DBSCAN(eps=eps, min_samples=min_samples, metric='cosine').fit(face_embeddings)
    
    # Group faces by cluster
    face_clusters = defaultdict(list)
    for i, label in enumerate(clustering.labels_):
        face_clusters[label].append(face_metadata[i])
    
    # Sort clusters by size
    sorted_clusters = sorted(face_clusters.items(), key=lambda x: len(x[1]), reverse=True)
    
    all_faces_summary = []
    prominent_faces = []
    
    for i, (label, cluster) in enumerate(sorted_clusters):
        if label != -1:  # Ignore noise points
            # Collect all appearances
            appearances = [
                {
                    'start': face['start'],
                    'end': face['end'],
                    'size_ratio': face.get('size_ratio', 1.0)  # Use 1.0 as default if size_ratio is not present
                }
                for face in cluster
            ]
            
            # Sort appearances by start time
            appearances.sort(key=lambda x: x['start'])
            
            # Select representative face (e.g., largest face in the cluster)
            representative_face = max(cluster, key=lambda f: f.get('size_ratio', 1.0))
            
            face_summary = {
                "id": f"face_{i}",
                "cluster_id": f"cluster_{label}",
                "bbox": representative_face.get('bbox', []),
                "embedding": representative_face.get('embedding', []),
                "appearances": appearances,
                "total_appearances": len(appearances),
                "total_screen_time": sum(app['end'] - app['start'] for app in appearances),
                "first_appearance": appearances[0]['start'],
                "last_appearance": appearances[-1]['end'],
                "thumbnail": representative_face.get('thumbnail', '')
            }
            
            all_faces_summary.append(face_summary)
            
            if i < top_k:
                prominent_faces.append(face_summary)
    
    return all_faces_summary, prominent_faces, face_embeddings, clustering.labels_

# Usage in the main Streamlit app:
all_faces_summary, prominent_faces, face_embeddings, face_labels = create_comprehensive_face_summary(face_index, face_metadata)

# Face cluster visualization
# Update the face cluster visualization function
def plot_face_clusters_interactive(face_embeddings, face_labels, all_faces_summary, prominent_faces):
    pca = PCA(n_components=3)
    embeddings_3d = pca.fit_transform(face_embeddings)
    
    unique_labels = set(face_labels)
    colors = [f'rgb({int(r*255)},{int(g*255)},{int(b*255)})' 
              for r, g, b, _ in plt.cm.rainbow(np.linspace(0, 1, len(unique_labels)))]
    
    traces = []
    for label, color in zip(unique_labels, colors):
        if label == -1:
            continue  # Skip noise points
        
        cluster_points = embeddings_3d[face_labels == label]
        cluster_faces = [face for face in all_faces_summary if face['cluster_id'] == f'cluster_{label}']
        
        hover_text = [
            f"Cluster {label}<br>"
            f"Time: {face['appearances'][0]['start']:.2f}s - {face['appearances'][-1]['end']:.2f}s<br>"
            f"Appearances: {face['total_appearances']}"
            for face in cluster_faces
        ]
        
        trace = go.Scatter3d(
            x=cluster_points[:, 0],
            y=cluster_points[:, 1],
            z=cluster_points[:, 2],
            mode='markers',
            name=f'Cluster {label}',
            marker=dict(size=5, color=color, opacity=0.8),
            text=hover_text,
            hoverinfo='text'
        )
        traces.append(trace)
    
    # Add markers for prominent faces
    prominent_points = [embeddings_3d[face_labels == int(face['cluster_id'].split('_')[1])][0] for face in prominent_faces]
    prominent_trace = go.Scatter3d(
        x=[p[0] for p in prominent_points],
        y=[p[1] for p in prominent_points],
        z=[p[2] for p in prominent_points],
        mode='markers',
        name='Prominent Faces',
        marker=dict(size=10, color='red', symbol='star'),
        text=[f"Prominent Face<br>Cluster {face['cluster_id']}" for face in prominent_faces],
        hoverinfo='text'
    )
    traces.append(prominent_trace)
    
    layout = go.Layout(
        title='Face Clusters Visualization',
        scene=dict(xaxis_title='PCA 1', yaxis_title='PCA 2', zaxis_title='PCA 3'),
        margin=dict(r=0, b=0, l=0, t=40)
    )
    
    fig = go.Figure(data=traces, layout=layout)
    return fig

# Search functions
def combined_search(query, text_index, image_index, text_metadata, image_metadata, text_model, image_model, n_results=5):
    if isinstance(query, str):
        text_vector = text_model.encode([query], convert_to_tensor=True).cpu().numpy()
        image_vector = image_model.encode([query], convert_to_tensor=True).cpu().numpy()
    else:  # Assume it's an image
        image_vector = image_model.encode(query, convert_to_tensor=True).cpu().numpy()
        text_vector = image_vector  # Use the same vector for text search in this case
    
    text_D, text_I = text_index.search(text_vector, n_results)
    image_D, image_I = image_index.search(image_vector, n_results)
    
    text_results = [{'data': text_metadata[i], 'distance': d, 'type': 'text'} for i, d in zip(text_I[0], text_D[0])]
    image_results = [{'data': image_metadata[i], 'distance': d, 'type': 'image'} for i, d in zip(image_I[0], image_D[0])]
    
    combined_results = sorted(text_results + image_results, key=lambda x: x['distance'])
    return combined_results[:n_results]

def face_search(face_embedding, face_index, face_metadata, n_results=5):
    D, I = face_index.search(np.array([face_embedding]), n_results)
    results = [face_metadata[i] for i in I[0]]
    return results, D[0]

def detect_and_embed_face(image, face_app):
    img_array = np.array(image)
    faces = face_app.get(img_array)
    if len(faces) == 0:
        return None
    largest_face = max(faces, key=lambda x: (x.bbox[2] - x.bbox[0]) * (x.bbox[3] - x.bbox[1]))
    return largest_face.embedding

def create_video_clip(video_path, start_time, end_time, output_path):
    with VideoFileClip(video_path) as video:
        new_clip = video.subclip(start_time, end_time)
        new_clip.write_videofile(output_path, codec="libx264", audio_codec="aac")
    return output_path

# Streamlit UI
st.title("Video Analysis Dashboard")

# Sidebar with full video and scrollable transcript
st.sidebar.header("Full Video")
st.sidebar.video(video_path)

st.sidebar.header("Video Transcript")
transcript_text = transcription['transcription']
st.sidebar.text_area("Full Transcript", transcript_text, height=300)

# Main content
st.header("Video Summary")

# Face Clusters
st.subheader("Prominent Face Clusters")
for face in prominent_faces:  # Use prominent_faces instead of face_summary
    st.write(f"Face Cluster {face['cluster_id']}:")
    st.write(f"  Total appearances: {face['total_appearances']}")
    st.write(f"  Total screen time: {face['total_screen_time']:.2f} seconds")
    st.write(f"  First appearance: {face['first_appearance']:.2f} seconds")
    st.write(f"  Last appearance: {face['last_appearance']:.2f} seconds")
    st.write(f"  Timeline: {len(face['appearances'])} appearances")
    st.write("  First 5 appearances:")
    for app in face['appearances'][:5]:
        st.write(f"    {app['start']:.2f}s - {app['end']:.2f}s")
    if face['thumbnail']:
        image = Image.open(io.BytesIO(base64.b64decode(face['thumbnail'])))
        st.image(image, caption=f"Representative face for {face['cluster_id']}", width=100)
    st.write("---")

# Face Cluster Visualization
st.subheader("Face Cluster Visualization")
fig = plot_face_clusters_interactive(face_embeddings, face_labels, all_faces_summary, prominent_faces)
st.plotly_chart(fig)

# Themes
st.subheader("Themes")
for theme in summary['themes']:
    st.write(f"Theme ID: {theme['id']}, Keywords: {', '.join(theme['keywords'])}")

# Search functionality
st.header("Search")

search_type = st.selectbox("Select search type", ["Combined", "Face"])

if search_type == "Combined":
    search_method = st.radio("Choose search method", ["Text", "Image"])
    
    if search_method == "Text":
        query = st.text_input("Enter your search query")
        if st.button("Search"):
            results = combined_search(query, text_index, image_index, text_metadata, image_metadata, text_model, image_model)
            st.subheader("Search Results")
            for result in results:
                st.write(f"Type: {result['type']}, Time: {result['data']['start']:.2f}s - {result['data']['end']:.2f}s, Distance: {result['distance']:.4f}")
                if 'text' in result['data']:
                    st.write(f"Text: {result['data']['text']}")
                clip_path = create_video_clip(video_path, result['data']['start'], result['data']['end'], f"temp_clip_{result['data']['start']}.mp4")
                st.video(clip_path)
                st.write("---")
    else:
        uploaded_file = st.file_uploader("Choose an image...", type=["jpg", "jpeg", "png"])
        if uploaded_file is not None:
            image = Image.open(uploaded_file)
            st.image(image, caption="Uploaded Image", use_column_width=True)
            if st.button("Search"):
                results = combined_search(image, text_index, image_index, text_metadata, image_metadata, text_model, image_model)
                st.subheader("Image Search Results")
                for result in results:
                    st.write(f"Type: {result['type']}, Time: {result['data']['start']:.2f}s - {result['data']['end']:.2f}s, Distance: {result['distance']:.4f}")
                    clip_path = create_video_clip(video_path, result['data']['start'], result['data']['end'], f"temp_clip_{result['data']['start']}.mp4")
                    st.video(clip_path)
                    st.write("---")

elif search_type == "Face":
    face_search_type = st.radio("Choose face search method", ["Select from clusters", "Upload image"])
    
    if face_search_type == "Select from clusters":
        cluster_id = st.selectbox("Select a face cluster", [cluster['cluster_id'] for cluster in face_summary])
        if st.button("Search"):
            selected_cluster = next(cluster for cluster in face_summary if cluster['cluster_id'] == cluster_id)
            st.subheader("Face Cluster Search Results")
            for appearance in selected_cluster['appearances'][:5]:  # Show first 5 appearances
                st.write(f"Time: {appearance['start']:.2f}s - {appearance['end']:.2f}s")
                clip_path = create_video_clip(video_path, appearance['start'], appearance['end'], f"temp_face_clip_{appearance['start']}.mp4")
                st.video(clip_path)
                st.write("---")
    else:
        uploaded_file = st.file_uploader("Choose a face image...", type=["jpg", "jpeg", "png"])
        if uploaded_file is not None:
            image = Image.open(uploaded_file)
            st.image(image, caption="Uploaded Image", use_column_width=True)
            if st.button("Search"):
                face_embedding = detect_and_embed_face(image, face_app)
                if face_embedding is not None:
                    face_results, face_distances = face_search(face_embedding, face_index, face_metadata)
                    st.subheader("Face Search Results")
                    for result, distance in zip(face_results, face_distances):
                        st.write(f"Time: {result['start']:.2f}s - {result['end']:.2f}s, Distance: {distance:.4f}")
                        clip_path = create_video_clip(video_path, result['start'], result['end'], f"temp_face_clip_{result['start']}.mp4")
                        st.video(clip_path)
                        st.write("---")
                else:
                    st.error("No face detected in the uploaded image. Please try another image.")