Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,401 +1,399 @@
|
|
| 1 |
-
import os
|
| 2 |
-
import time
|
| 3 |
-
import cv2
|
| 4 |
-
import numpy as np
|
| 5 |
-
import tensorflow as tf
|
| 6 |
-
import gradio as gr
|
| 7 |
-
import plotly.graph_objects as go
|
| 8 |
-
import matplotlib.pyplot as plt
|
| 9 |
-
from fpdf import FPDF
|
| 10 |
-
from PIL import Image
|
| 11 |
-
|
| 12 |
-
# ===============================
|
| 13 |
-
# 1. Load Model
|
| 14 |
-
# ===============================
|
| 15 |
-
MODEL_PATH = "fer_surprise_softmax.h5"
|
| 16 |
-
model = tf.keras.models.load_model(MODEL_PATH, compile=False)
|
| 17 |
-
|
| 18 |
-
IMG_SIZE = (96, 96)
|
| 19 |
-
CLASS_NAMES = ["angry", "disgust", "fear", "happy", "sad", "surprise", "neutral"]
|
| 20 |
-
SURPRISE_IDX = CLASS_NAMES.index("surprise")
|
| 21 |
-
|
| 22 |
-
# ===============================
|
| 23 |
-
# 2. Face Detector
|
| 24 |
-
# ===============================
|
| 25 |
-
face_cascade = cv2.CascadeClassifier(
|
| 26 |
-
cv2.data.haarcascades + "haarcascade_frontalface_default.xml"
|
| 27 |
-
)
|
| 28 |
-
|
| 29 |
-
# ===============================
|
| 30 |
-
# 3. State Storage
|
| 31 |
-
# ===============================
|
| 32 |
-
events = []
|
| 33 |
-
surprise_history = []
|
| 34 |
-
start_time = None
|
| 35 |
-
MIN_EVENT_GAP = 1.0
|
| 36 |
-
|
| 37 |
-
# Session stats
|
| 38 |
-
frames_with_face = 0
|
| 39 |
-
max_p_surprise = 0.0
|
| 40 |
-
|
| 41 |
-
# ===============================
|
| 42 |
-
# 4. Utility: Time Formatting
|
| 43 |
-
# ===============================
|
| 44 |
-
def format_time(seconds: float) -> str:
|
| 45 |
-
minutes = int(seconds // 60)
|
| 46 |
-
sec = int(seconds % 60)
|
| 47 |
-
return f"{minutes:02d}:{sec:02d}"
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
# ===============================
|
| 51 |
-
# 5. Real-time Frame Processing
|
| 52 |
-
# ===============================
|
| 53 |
-
def detect_surprise(frame, threshold):
|
| 54 |
-
|
| 55 |
-
global events, start_time, surprise_history
|
| 56 |
-
global frames_with_face, max_p_surprise
|
| 57 |
-
|
| 58 |
-
if frame is None:
|
| 59 |
-
stats_text = (
|
| 60 |
-
"### Session Stats\n"
|
| 61 |
-
"- Session duration: 00:00\n"
|
| 62 |
-
f"- Current threshold: {threshold:.2f}\n"
|
| 63 |
-
"- Frames with face detected: 0\n"
|
| 64 |
-
"- Surprise events detected: 0\n"
|
| 65 |
-
"- Max P(surprise): 0.00\n"
|
| 66 |
-
)
|
| 67 |
-
return None, {"Error": 1.0}, None, stats_text
|
| 68 |
-
|
| 69 |
-
if start_time is None:
|
| 70 |
-
start_time = time.time()
|
| 71 |
-
surprise_history = []
|
| 72 |
-
events = []
|
| 73 |
-
frames_with_face = 0
|
| 74 |
-
max_p_surprise = 0.0
|
| 75 |
-
|
| 76 |
-
current_time = time.time() - start_time
|
| 77 |
-
|
| 78 |
-
frame_bgr = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
|
| 79 |
-
gray = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2GRAY)
|
| 80 |
-
|
| 81 |
-
faces = face_cascade.detectMultiScale(gray, 1.1, 4)
|
| 82 |
-
|
| 83 |
-
# 변경된 기본 라벨: 얼굴 미검출 시 조명/각도 안내
|
| 84 |
-
label = "NO FACE - Try brighter lighting or adjust angle"
|
| 85 |
-
color = (0, 255, 255)
|
| 86 |
-
probs_dict = {}
|
| 87 |
-
|
| 88 |
-
if len(faces) > 0:
|
| 89 |
-
frames_with_face += 1
|
| 90 |
-
x, y, w, h = sorted(faces, key=lambda r: r[2] * r[3], reverse=True)[0]
|
| 91 |
-
roi = frame_bgr[y:y+h, x:x+w]
|
| 92 |
-
|
| 93 |
-
rgb = cv2.cvtColor(roi, cv2.COLOR_BGR2RGB)
|
| 94 |
-
resized = cv2.resize(rgb, IMG_SIZE)
|
| 95 |
-
inp = resized.astype("float32") / 255.0
|
| 96 |
-
inp = np.expand_dims(inp, axis=0)
|
| 97 |
-
|
| 98 |
-
probs = model.predict(inp, verbose=0)[0]
|
| 99 |
-
p_surprise = float(probs[SURPRISE_IDX])
|
| 100 |
-
|
| 101 |
-
if p_surprise > max_p_surprise:
|
| 102 |
-
max_p_surprise = p_surprise
|
| 103 |
-
|
| 104 |
-
probs_dict = {
|
| 105 |
-
cls: float(p) for cls, p in zip(CLASS_NAMES, probs)
|
| 106 |
-
}
|
| 107 |
-
|
| 108 |
-
surprise_history.append({
|
| 109 |
-
"time": current_time,
|
| 110 |
-
"score": p_surprise,
|
| 111 |
-
})
|
| 112 |
-
|
| 113 |
-
# -------- Top3 detection logic --------
|
| 114 |
-
if p_surprise >= threshold:
|
| 115 |
-
if len(events) == 0:
|
| 116 |
-
events.append({
|
| 117 |
-
"time": current_time,
|
| 118 |
-
"score": p_surprise,
|
| 119 |
-
"frame": frame.copy()
|
| 120 |
-
})
|
| 121 |
-
else:
|
| 122 |
-
dt = current_time - events[-1]["time"]
|
| 123 |
-
if dt > MIN_EVENT_GAP:
|
| 124 |
-
events.append({
|
| 125 |
-
"time": current_time,
|
| 126 |
-
"score": p_surprise,
|
| 127 |
-
"frame": frame.copy()
|
| 128 |
-
})
|
| 129 |
-
else:
|
| 130 |
-
if p_surprise > events[-1]["score"]:
|
| 131 |
-
events[-1]["time"] = current_time
|
| 132 |
-
events[-1]["score"] = p_surprise
|
| 133 |
-
events[-1]["frame"] = frame.copy()
|
| 134 |
-
|
| 135 |
-
label = f"😲 SURPRISE (p={p_surprise:.2f})"
|
| 136 |
-
color = (0, 255, 0)
|
| 137 |
-
|
| 138 |
-
else:
|
| 139 |
-
label = f"🙂 Not Surprise (p={p_surprise:.2f})"
|
| 140 |
-
color = (0, 0, 255)
|
| 141 |
-
|
| 142 |
-
# Draw bounding box
|
| 143 |
-
cv2.rectangle(frame_bgr, (x, y), (x + w, y + h), color, 3)
|
| 144 |
-
|
| 145 |
-
# -------- Label 위치: 왼쪽 아래 + 큰 글씨 --------
|
| 146 |
-
h_img, w_img = frame_bgr.shape[:2]
|
| 147 |
-
cv2.putText(
|
| 148 |
-
frame_bgr,
|
| 149 |
-
label,
|
| 150 |
-
(10, h_img - 10),
|
| 151 |
-
cv2.FONT_HERSHEY_SIMPLEX,
|
| 152 |
-
1.6,
|
| 153 |
-
color,
|
| 154 |
-
3
|
| 155 |
-
)
|
| 156 |
-
|
| 157 |
-
out_rgb = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB)
|
| 158 |
-
|
| 159 |
-
# Per-frame bar chart
|
| 160 |
-
fig = go.Figure()
|
| 161 |
-
if len(probs_dict) > 0:
|
| 162 |
-
fig.add_trace(go.Bar(
|
| 163 |
-
x=list(probs_dict.keys()),
|
| 164 |
-
y=list(probs_dict.values()),
|
| 165 |
-
marker_color="lightskyblue"
|
| 166 |
-
))
|
| 167 |
-
fig.update_layout(
|
| 168 |
-
title="Emotion Probability Distribution",
|
| 169 |
-
yaxis=dict(range=[0, 1])
|
| 170 |
-
)
|
| 171 |
-
|
| 172 |
-
session_duration_str = format_time(current_time)
|
| 173 |
-
stats_text = (
|
| 174 |
-
"### Session Stats\n"
|
| 175 |
-
f"- Session duration: {session_duration_str}\n"
|
| 176 |
-
f"- Current threshold: {threshold:.2f}\n"
|
| 177 |
-
f"- Frames with face detected: {frames_with_face}\n"
|
| 178 |
-
f"- Surprise events detected: {len(events)}\n"
|
| 179 |
-
f"- Max P(surprise): {max_p_surprise:.2f}\n"
|
| 180 |
-
)
|
| 181 |
-
|
| 182 |
-
return out_rgb, probs_dict, fig, stats_text
|
| 183 |
-
|
| 184 |
-
|
| 185 |
-
# ===============================
|
| 186 |
-
# 6. PDF Generation
|
| 187 |
-
# ===============================
|
| 188 |
-
def create_pdf(summary_text, top_images, timeline_fig):
|
| 189 |
-
os.makedirs("reports", exist_ok=True)
|
| 190 |
-
timestamp = int(time.time())
|
| 191 |
-
pdf_path = os.path.join("reports", f"surprise_report_{timestamp}.pdf")
|
| 192 |
-
|
| 193 |
-
timeline_path = os.path.join("reports", f"timeline_{timestamp}.png")
|
| 194 |
-
timeline_fig.savefig(timeline_path, bbox_inches="tight")
|
| 195 |
-
|
| 196 |
-
img_paths = []
|
| 197 |
-
for i, img in enumerate(top_images):
|
| 198 |
-
if img is None:
|
| 199 |
-
img_paths.append(None)
|
| 200 |
-
continue
|
| 201 |
-
img_pil = Image.fromarray(img)
|
| 202 |
-
img_path = os.path.join("reports", f"top{i+1}_{timestamp}.png")
|
| 203 |
-
img_pil.save(img_path)
|
| 204 |
-
img_paths.append(img_path)
|
| 205 |
-
|
| 206 |
-
pdf = FPDF()
|
| 207 |
-
pdf.add_page()
|
| 208 |
-
|
| 209 |
-
pdf.set_font("Arial", "B", 16)
|
| 210 |
-
pdf.cell(0, 10, "Real-Time Surprise Detector Report", ln=1)
|
| 211 |
-
|
| 212 |
-
pdf.set_font("Arial", "", 11)
|
| 213 |
-
pdf.multi_cell(0, 6, summary_text)
|
| 214 |
-
pdf.ln(4)
|
| 215 |
-
|
| 216 |
-
pdf.set_font("Arial", "B", 12)
|
| 217 |
-
pdf.cell(0, 8, "Surprise Probability Timeline", ln=1)
|
| 218 |
-
pdf.image(timeline_path, w=170)
|
| 219 |
-
pdf.ln(4)
|
| 220 |
-
|
| 221 |
-
pdf.set_font("Arial", "B", 12)
|
| 222 |
-
pdf.cell(0, 8, "Top Surprise Frames", ln=1)
|
| 223 |
-
pdf.set_font("Arial", "", 11)
|
| 224 |
-
|
| 225 |
-
for i, path in enumerate(img_paths):
|
| 226 |
-
if path is not None:
|
| 227 |
-
pdf.cell(0, 6, f"Top {i+1}", ln=1)
|
| 228 |
-
pdf.image(path, w=80)
|
| 229 |
-
pdf.ln(2)
|
| 230 |
-
|
| 231 |
-
pdf.output(pdf_path)
|
| 232 |
-
return pdf_path
|
| 233 |
-
|
| 234 |
-
|
| 235 |
-
# ===============================
|
| 236 |
-
# 7. Summarize Results
|
| 237 |
-
# ===============================
|
| 238 |
-
def summarize_results():
|
| 239 |
-
|
| 240 |
-
global events, start_time, surprise_history
|
| 241 |
-
global frames_with_face, max_p_surprise
|
| 242 |
-
|
| 243 |
-
if len(surprise_history) == 0:
|
| 244 |
-
return "No data recorded.", None, None, None, None, None
|
| 245 |
-
|
| 246 |
-
times = [h["time"] for h in surprise_history]
|
| 247 |
-
scores = [h["score"] for h in surprise_history]
|
| 248 |
-
|
| 249 |
-
fig, ax = plt.subplots()
|
| 250 |
-
ax.plot(times, scores, marker="o", linewidth=1)
|
| 251 |
-
ax.set_title("Surprise Probability Timeline")
|
| 252 |
-
ax.set_xlabel("Time (s)")
|
| 253 |
-
ax.set_ylabel("P(surprise)")
|
| 254 |
-
ax.set_ylim(0, 1)
|
| 255 |
-
ax.grid(True)
|
| 256 |
-
|
| 257 |
-
top_images = [None, None, None]
|
| 258 |
-
if len(events) == 0:
|
| 259 |
-
summary_text = (
|
| 260 |
-
"No surprise events detected above the current threshold.\n\n"
|
| 261 |
-
"The timeline shows overall surprise probability over time."
|
| 262 |
-
)
|
| 263 |
-
img1 = img2 = img3 = None
|
| 264 |
-
|
| 265 |
-
else:
|
| 266 |
-
top3 = sorted(events, key=lambda x: x["score"], reverse=True)[:3]
|
| 267 |
-
|
| 268 |
-
captions = []
|
| 269 |
-
images = []
|
| 270 |
-
top_times = []
|
| 271 |
-
top_scores = []
|
| 272 |
-
|
| 273 |
-
for i, e in enumerate(top3):
|
| 274 |
-
formatted_time = format_time(e["time"])
|
| 275 |
-
score = e["score"]
|
| 276 |
-
captions.append(f"#{i+1} Time = {formatted_time} Score = {score:.2f}")
|
| 277 |
-
images.append(e["frame"])
|
| 278 |
-
top_times.append(e["time"])
|
| 279 |
-
top_scores.append(score)
|
| 280 |
-
|
| 281 |
-
summary_text = "Top 3 surprise moments:\n" + "\n".join(captions)
|
| 282 |
-
|
| 283 |
-
markers = ["*", "^", "s"]
|
| 284 |
-
colors = ["red", "darkorange", "gold"]
|
| 285 |
-
|
| 286 |
-
for i, (t, s) in enumerate(zip(top_times, top_scores)):
|
| 287 |
-
ax.scatter(t, s, color=colors[i], marker=markers[i], s=80, zorder=5)
|
| 288 |
-
|
| 289 |
-
for i in range(3):
|
| 290 |
-
if i < len(images):
|
| 291 |
-
top_images[i] = images[i]
|
| 292 |
-
|
| 293 |
-
img1, img2, img3 = top_images
|
| 294 |
-
|
| 295 |
-
pdf_path = create_pdf(summary_text, top_images, fig)
|
| 296 |
-
|
| 297 |
-
events = []
|
| 298 |
-
start_time = None
|
| 299 |
-
surprise_history = []
|
| 300 |
-
frames_with_face = 0
|
| 301 |
-
max_p_surprise = 0.0
|
| 302 |
-
|
| 303 |
-
return summary_text, img1, img2, img3, fig, pdf_path
|
| 304 |
-
|
| 305 |
-
|
| 306 |
-
# ===============================
|
| 307 |
-
# 8. UI
|
| 308 |
-
# ===============================
|
| 309 |
-
try:
|
| 310 |
-
custom_theme = gr.themes.Soft(primary_hue="indigo", neutral_hue="slate")
|
| 311 |
-
except:
|
| 312 |
-
custom_theme = "soft"
|
| 313 |
-
|
| 314 |
-
demo = gr.Blocks(theme=custom_theme)
|
| 315 |
-
|
| 316 |
-
with demo:
|
| 317 |
-
|
| 318 |
-
gr.Markdown(
|
| 319 |
-
"""
|
| 320 |
-
# 🎭 Real-Time Surprise Detector
|
| 321 |
-
### A real-time facial reaction analysis system
|
| 322 |
-
##### Detects surprise reactions using facial emotion recognition and summarizes top 3 peak surprise moments.
|
| 323 |
-
|
| 324 |
-
**How to use:**
|
| 325 |
-
1. Enable your webcam by clicking the feed area.
|
| 326 |
-
2. Watch your chosen video while keeping your face visible.
|
| 327 |
-
3. If many frames show **"NO FACE"**, try brighter lighting or adjust your face angle.
|
| 328 |
-
4. Click **“Show Top 3 Surprise Moments”** after stopping the stream.
|
| 329 |
-
5. Download the generated PDF if needed.
|
| 330 |
-
---
|
| 331 |
-
"""
|
| 332 |
-
)
|
| 333 |
-
|
| 334 |
-
with gr.Row():
|
| 335 |
-
with gr.Column(scale=2):
|
| 336 |
-
|
| 337 |
-
webcam = gr.Image(
|
| 338 |
-
sources=["webcam"],
|
| 339 |
-
type="numpy",
|
| 340 |
-
label="Webcam Feed"
|
| 341 |
-
)
|
| 342 |
-
output_img = gr.Image(label="Detection Result")
|
| 343 |
-
|
| 344 |
-
with gr.Column(scale=1):
|
| 345 |
-
threshold = gr.Slider(
|
| 346 |
-
minimum=0.0, maximum=1.0, value=0.1,
|
| 347 |
-
step=0.01, label="Surprise Threshold"
|
| 348 |
-
)
|
| 349 |
-
|
| 350 |
-
gr.Markdown(
|
| 351 |
-
"""
|
| 352 |
-
### What is the Surprise Threshold?
|
| 353 |
-
|
| 354 |
-
- Lower threshold → detects smaller reactions
|
| 355 |
-
- Higher threshold → detects only strong surprise
|
| 356 |
-
- **Default = 0.1**
|
| 357 |
-
|
| 358 |
-
👉 Try making a surprised face to adjust sensitivity.
|
| 359 |
-
"""
|
| 360 |
-
)
|
| 361 |
-
|
| 362 |
-
output_label = gr.Label(label="Softmax Probabilities")
|
| 363 |
-
plot = gr.Plot(label="Emotion Probability (per frame)")
|
| 364 |
-
stats_md = gr.Markdown("### Session Stats\nWaiting for stream...")
|
| 365 |
-
|
| 366 |
-
webcam.stream(
|
| 367 |
-
fn=detect_surprise,
|
| 368 |
-
inputs=[webcam, threshold],
|
| 369 |
-
outputs=[output_img, output_label, plot, stats_md],
|
| 370 |
-
stream_every=0.1
|
| 371 |
-
)
|
| 372 |
-
|
| 373 |
-
gr.Markdown("---")
|
| 374 |
-
gr.Markdown("## 🔍 Summary & Report")
|
| 375 |
-
|
| 376 |
-
summarize_button = gr.Button("🎯 Show Top 3 Surprise Moments")
|
| 377 |
-
|
| 378 |
-
summary_text = gr.Textbox(
|
| 379 |
-
label="Top 3 Summary",
|
| 380 |
-
lines=6,
|
| 381 |
-
max_lines=10
|
| 382 |
-
)
|
| 383 |
-
|
| 384 |
-
with gr.Row():
|
| 385 |
-
img1 = gr.Image(label="Top 1")
|
| 386 |
-
img2 = gr.Image(label="Top 2")
|
| 387 |
-
img3 = gr.Image(label="Top 3")
|
| 388 |
-
|
| 389 |
-
timeline_plot = gr.Plot(label="Surprise Timeline")
|
| 390 |
-
pdf_file = gr.File(label="Download PDF Report")
|
| 391 |
-
|
| 392 |
-
summarize_button.click(
|
| 393 |
-
fn=summarize_results,
|
| 394 |
-
inputs=[],
|
| 395 |
-
outputs=[summary_text, img1, img2, img3, timeline_plot, pdf_file]
|
| 396 |
-
)
|
| 397 |
-
|
| 398 |
-
if __name__ == "__main__":
|
| 399 |
-
demo.launch()
|
| 400 |
-
|
| 401 |
-
::contentReference[oaicite:0]{index=0}
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import time
|
| 3 |
+
import cv2
|
| 4 |
+
import numpy as np
|
| 5 |
+
import tensorflow as tf
|
| 6 |
+
import gradio as gr
|
| 7 |
+
import plotly.graph_objects as go
|
| 8 |
+
import matplotlib.pyplot as plt
|
| 9 |
+
from fpdf import FPDF
|
| 10 |
+
from PIL import Image
|
| 11 |
+
|
| 12 |
+
# ===============================
|
| 13 |
+
# 1. Load Model
|
| 14 |
+
# ===============================
|
| 15 |
+
MODEL_PATH = "fer_surprise_softmax.h5"
|
| 16 |
+
model = tf.keras.models.load_model(MODEL_PATH, compile=False)
|
| 17 |
+
|
| 18 |
+
IMG_SIZE = (96, 96)
|
| 19 |
+
CLASS_NAMES = ["angry", "disgust", "fear", "happy", "sad", "surprise", "neutral"]
|
| 20 |
+
SURPRISE_IDX = CLASS_NAMES.index("surprise")
|
| 21 |
+
|
| 22 |
+
# ===============================
|
| 23 |
+
# 2. Face Detector
|
| 24 |
+
# ===============================
|
| 25 |
+
face_cascade = cv2.CascadeClassifier(
|
| 26 |
+
cv2.data.haarcascades + "haarcascade_frontalface_default.xml"
|
| 27 |
+
)
|
| 28 |
+
|
| 29 |
+
# ===============================
|
| 30 |
+
# 3. State Storage
|
| 31 |
+
# ===============================
|
| 32 |
+
events = []
|
| 33 |
+
surprise_history = []
|
| 34 |
+
start_time = None
|
| 35 |
+
MIN_EVENT_GAP = 1.0
|
| 36 |
+
|
| 37 |
+
# Session stats
|
| 38 |
+
frames_with_face = 0
|
| 39 |
+
max_p_surprise = 0.0
|
| 40 |
+
|
| 41 |
+
# ===============================
|
| 42 |
+
# 4. Utility: Time Formatting
|
| 43 |
+
# ===============================
|
| 44 |
+
def format_time(seconds: float) -> str:
|
| 45 |
+
minutes = int(seconds // 60)
|
| 46 |
+
sec = int(seconds % 60)
|
| 47 |
+
return f"{minutes:02d}:{sec:02d}"
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
# ===============================
|
| 51 |
+
# 5. Real-time Frame Processing
|
| 52 |
+
# ===============================
|
| 53 |
+
def detect_surprise(frame, threshold):
|
| 54 |
+
|
| 55 |
+
global events, start_time, surprise_history
|
| 56 |
+
global frames_with_face, max_p_surprise
|
| 57 |
+
|
| 58 |
+
if frame is None:
|
| 59 |
+
stats_text = (
|
| 60 |
+
"### Session Stats\n"
|
| 61 |
+
"- Session duration: 00:00\n"
|
| 62 |
+
f"- Current threshold: {threshold:.2f}\n"
|
| 63 |
+
"- Frames with face detected: 0\n"
|
| 64 |
+
"- Surprise events detected: 0\n"
|
| 65 |
+
"- Max P(surprise): 0.00\n"
|
| 66 |
+
)
|
| 67 |
+
return None, {"Error": 1.0}, None, stats_text
|
| 68 |
+
|
| 69 |
+
if start_time is None:
|
| 70 |
+
start_time = time.time()
|
| 71 |
+
surprise_history = []
|
| 72 |
+
events = []
|
| 73 |
+
frames_with_face = 0
|
| 74 |
+
max_p_surprise = 0.0
|
| 75 |
+
|
| 76 |
+
current_time = time.time() - start_time
|
| 77 |
+
|
| 78 |
+
frame_bgr = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
|
| 79 |
+
gray = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2GRAY)
|
| 80 |
+
|
| 81 |
+
faces = face_cascade.detectMultiScale(gray, 1.1, 4)
|
| 82 |
+
|
| 83 |
+
# 변경된 기본 라벨: 얼굴 미검출 시 조명/각도 안내
|
| 84 |
+
label = "NO FACE - Try brighter lighting or adjust angle"
|
| 85 |
+
color = (0, 255, 255)
|
| 86 |
+
probs_dict = {}
|
| 87 |
+
|
| 88 |
+
if len(faces) > 0:
|
| 89 |
+
frames_with_face += 1
|
| 90 |
+
x, y, w, h = sorted(faces, key=lambda r: r[2] * r[3], reverse=True)[0]
|
| 91 |
+
roi = frame_bgr[y:y+h, x:x+w]
|
| 92 |
+
|
| 93 |
+
rgb = cv2.cvtColor(roi, cv2.COLOR_BGR2RGB)
|
| 94 |
+
resized = cv2.resize(rgb, IMG_SIZE)
|
| 95 |
+
inp = resized.astype("float32") / 255.0
|
| 96 |
+
inp = np.expand_dims(inp, axis=0)
|
| 97 |
+
|
| 98 |
+
probs = model.predict(inp, verbose=0)[0]
|
| 99 |
+
p_surprise = float(probs[SURPRISE_IDX])
|
| 100 |
+
|
| 101 |
+
if p_surprise > max_p_surprise:
|
| 102 |
+
max_p_surprise = p_surprise
|
| 103 |
+
|
| 104 |
+
probs_dict = {
|
| 105 |
+
cls: float(p) for cls, p in zip(CLASS_NAMES, probs)
|
| 106 |
+
}
|
| 107 |
+
|
| 108 |
+
surprise_history.append({
|
| 109 |
+
"time": current_time,
|
| 110 |
+
"score": p_surprise,
|
| 111 |
+
})
|
| 112 |
+
|
| 113 |
+
# -------- Top3 detection logic --------
|
| 114 |
+
if p_surprise >= threshold:
|
| 115 |
+
if len(events) == 0:
|
| 116 |
+
events.append({
|
| 117 |
+
"time": current_time,
|
| 118 |
+
"score": p_surprise,
|
| 119 |
+
"frame": frame.copy()
|
| 120 |
+
})
|
| 121 |
+
else:
|
| 122 |
+
dt = current_time - events[-1]["time"]
|
| 123 |
+
if dt > MIN_EVENT_GAP:
|
| 124 |
+
events.append({
|
| 125 |
+
"time": current_time,
|
| 126 |
+
"score": p_surprise,
|
| 127 |
+
"frame": frame.copy()
|
| 128 |
+
})
|
| 129 |
+
else:
|
| 130 |
+
if p_surprise > events[-1]["score"]:
|
| 131 |
+
events[-1]["time"] = current_time
|
| 132 |
+
events[-1]["score"] = p_surprise
|
| 133 |
+
events[-1]["frame"] = frame.copy()
|
| 134 |
+
|
| 135 |
+
label = f"😲 SURPRISE (p={p_surprise:.2f})"
|
| 136 |
+
color = (0, 255, 0)
|
| 137 |
+
|
| 138 |
+
else:
|
| 139 |
+
label = f"🙂 Not Surprise (p={p_surprise:.2f})"
|
| 140 |
+
color = (0, 0, 255)
|
| 141 |
+
|
| 142 |
+
# Draw bounding box
|
| 143 |
+
cv2.rectangle(frame_bgr, (x, y), (x + w, y + h), color, 3)
|
| 144 |
+
|
| 145 |
+
# -------- Label 위치: 왼쪽 아래 + 큰 글씨 --------
|
| 146 |
+
h_img, w_img = frame_bgr.shape[:2]
|
| 147 |
+
cv2.putText(
|
| 148 |
+
frame_bgr,
|
| 149 |
+
label,
|
| 150 |
+
(10, h_img - 10),
|
| 151 |
+
cv2.FONT_HERSHEY_SIMPLEX,
|
| 152 |
+
1.6,
|
| 153 |
+
color,
|
| 154 |
+
3
|
| 155 |
+
)
|
| 156 |
+
|
| 157 |
+
out_rgb = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB)
|
| 158 |
+
|
| 159 |
+
# Per-frame bar chart
|
| 160 |
+
fig = go.Figure()
|
| 161 |
+
if len(probs_dict) > 0:
|
| 162 |
+
fig.add_trace(go.Bar(
|
| 163 |
+
x=list(probs_dict.keys()),
|
| 164 |
+
y=list(probs_dict.values()),
|
| 165 |
+
marker_color="lightskyblue"
|
| 166 |
+
))
|
| 167 |
+
fig.update_layout(
|
| 168 |
+
title="Emotion Probability Distribution",
|
| 169 |
+
yaxis=dict(range=[0, 1])
|
| 170 |
+
)
|
| 171 |
+
|
| 172 |
+
session_duration_str = format_time(current_time)
|
| 173 |
+
stats_text = (
|
| 174 |
+
"### Session Stats\n"
|
| 175 |
+
f"- Session duration: {session_duration_str}\n"
|
| 176 |
+
f"- Current threshold: {threshold:.2f}\n"
|
| 177 |
+
f"- Frames with face detected: {frames_with_face}\n"
|
| 178 |
+
f"- Surprise events detected: {len(events)}\n"
|
| 179 |
+
f"- Max P(surprise): {max_p_surprise:.2f}\n"
|
| 180 |
+
)
|
| 181 |
+
|
| 182 |
+
return out_rgb, probs_dict, fig, stats_text
|
| 183 |
+
|
| 184 |
+
|
| 185 |
+
# ===============================
|
| 186 |
+
# 6. PDF Generation
|
| 187 |
+
# ===============================
|
| 188 |
+
def create_pdf(summary_text, top_images, timeline_fig):
|
| 189 |
+
os.makedirs("reports", exist_ok=True)
|
| 190 |
+
timestamp = int(time.time())
|
| 191 |
+
pdf_path = os.path.join("reports", f"surprise_report_{timestamp}.pdf")
|
| 192 |
+
|
| 193 |
+
timeline_path = os.path.join("reports", f"timeline_{timestamp}.png")
|
| 194 |
+
timeline_fig.savefig(timeline_path, bbox_inches="tight")
|
| 195 |
+
|
| 196 |
+
img_paths = []
|
| 197 |
+
for i, img in enumerate(top_images):
|
| 198 |
+
if img is None:
|
| 199 |
+
img_paths.append(None)
|
| 200 |
+
continue
|
| 201 |
+
img_pil = Image.fromarray(img)
|
| 202 |
+
img_path = os.path.join("reports", f"top{i+1}_{timestamp}.png")
|
| 203 |
+
img_pil.save(img_path)
|
| 204 |
+
img_paths.append(img_path)
|
| 205 |
+
|
| 206 |
+
pdf = FPDF()
|
| 207 |
+
pdf.add_page()
|
| 208 |
+
|
| 209 |
+
pdf.set_font("Arial", "B", 16)
|
| 210 |
+
pdf.cell(0, 10, "Real-Time Surprise Detector Report", ln=1)
|
| 211 |
+
|
| 212 |
+
pdf.set_font("Arial", "", 11)
|
| 213 |
+
pdf.multi_cell(0, 6, summary_text)
|
| 214 |
+
pdf.ln(4)
|
| 215 |
+
|
| 216 |
+
pdf.set_font("Arial", "B", 12)
|
| 217 |
+
pdf.cell(0, 8, "Surprise Probability Timeline", ln=1)
|
| 218 |
+
pdf.image(timeline_path, w=170)
|
| 219 |
+
pdf.ln(4)
|
| 220 |
+
|
| 221 |
+
pdf.set_font("Arial", "B", 12)
|
| 222 |
+
pdf.cell(0, 8, "Top Surprise Frames", ln=1)
|
| 223 |
+
pdf.set_font("Arial", "", 11)
|
| 224 |
+
|
| 225 |
+
for i, path in enumerate(img_paths):
|
| 226 |
+
if path is not None:
|
| 227 |
+
pdf.cell(0, 6, f"Top {i+1}", ln=1)
|
| 228 |
+
pdf.image(path, w=80)
|
| 229 |
+
pdf.ln(2)
|
| 230 |
+
|
| 231 |
+
pdf.output(pdf_path)
|
| 232 |
+
return pdf_path
|
| 233 |
+
|
| 234 |
+
|
| 235 |
+
# ===============================
|
| 236 |
+
# 7. Summarize Results
|
| 237 |
+
# ===============================
|
| 238 |
+
def summarize_results():
|
| 239 |
+
|
| 240 |
+
global events, start_time, surprise_history
|
| 241 |
+
global frames_with_face, max_p_surprise
|
| 242 |
+
|
| 243 |
+
if len(surprise_history) == 0:
|
| 244 |
+
return "No data recorded.", None, None, None, None, None
|
| 245 |
+
|
| 246 |
+
times = [h["time"] for h in surprise_history]
|
| 247 |
+
scores = [h["score"] for h in surprise_history]
|
| 248 |
+
|
| 249 |
+
fig, ax = plt.subplots()
|
| 250 |
+
ax.plot(times, scores, marker="o", linewidth=1)
|
| 251 |
+
ax.set_title("Surprise Probability Timeline")
|
| 252 |
+
ax.set_xlabel("Time (s)")
|
| 253 |
+
ax.set_ylabel("P(surprise)")
|
| 254 |
+
ax.set_ylim(0, 1)
|
| 255 |
+
ax.grid(True)
|
| 256 |
+
|
| 257 |
+
top_images = [None, None, None]
|
| 258 |
+
if len(events) == 0:
|
| 259 |
+
summary_text = (
|
| 260 |
+
"No surprise events detected above the current threshold.\n\n"
|
| 261 |
+
"The timeline shows overall surprise probability over time."
|
| 262 |
+
)
|
| 263 |
+
img1 = img2 = img3 = None
|
| 264 |
+
|
| 265 |
+
else:
|
| 266 |
+
top3 = sorted(events, key=lambda x: x["score"], reverse=True)[:3]
|
| 267 |
+
|
| 268 |
+
captions = []
|
| 269 |
+
images = []
|
| 270 |
+
top_times = []
|
| 271 |
+
top_scores = []
|
| 272 |
+
|
| 273 |
+
for i, e in enumerate(top3):
|
| 274 |
+
formatted_time = format_time(e["time"])
|
| 275 |
+
score = e["score"]
|
| 276 |
+
captions.append(f"#{i+1} Time = {formatted_time} Score = {score:.2f}")
|
| 277 |
+
images.append(e["frame"])
|
| 278 |
+
top_times.append(e["time"])
|
| 279 |
+
top_scores.append(score)
|
| 280 |
+
|
| 281 |
+
summary_text = "Top 3 surprise moments:\n" + "\n".join(captions)
|
| 282 |
+
|
| 283 |
+
markers = ["*", "^", "s"]
|
| 284 |
+
colors = ["red", "darkorange", "gold"]
|
| 285 |
+
|
| 286 |
+
for i, (t, s) in enumerate(zip(top_times, top_scores)):
|
| 287 |
+
ax.scatter(t, s, color=colors[i], marker=markers[i], s=80, zorder=5)
|
| 288 |
+
|
| 289 |
+
for i in range(3):
|
| 290 |
+
if i < len(images):
|
| 291 |
+
top_images[i] = images[i]
|
| 292 |
+
|
| 293 |
+
img1, img2, img3 = top_images
|
| 294 |
+
|
| 295 |
+
pdf_path = create_pdf(summary_text, top_images, fig)
|
| 296 |
+
|
| 297 |
+
events = []
|
| 298 |
+
start_time = None
|
| 299 |
+
surprise_history = []
|
| 300 |
+
frames_with_face = 0
|
| 301 |
+
max_p_surprise = 0.0
|
| 302 |
+
|
| 303 |
+
return summary_text, img1, img2, img3, fig, pdf_path
|
| 304 |
+
|
| 305 |
+
|
| 306 |
+
# ===============================
|
| 307 |
+
# 8. UI
|
| 308 |
+
# ===============================
|
| 309 |
+
try:
|
| 310 |
+
custom_theme = gr.themes.Soft(primary_hue="indigo", neutral_hue="slate")
|
| 311 |
+
except:
|
| 312 |
+
custom_theme = "soft"
|
| 313 |
+
|
| 314 |
+
demo = gr.Blocks(theme=custom_theme)
|
| 315 |
+
|
| 316 |
+
with demo:
|
| 317 |
+
|
| 318 |
+
gr.Markdown(
|
| 319 |
+
"""
|
| 320 |
+
# 🎭 Real-Time Surprise Detector
|
| 321 |
+
### A real-time facial reaction analysis system
|
| 322 |
+
##### Detects surprise reactions using facial emotion recognition and summarizes top 3 peak surprise moments.
|
| 323 |
+
|
| 324 |
+
**How to use:**
|
| 325 |
+
1. Enable your webcam by clicking the feed area.
|
| 326 |
+
2. Watch your chosen video while keeping your face visible.
|
| 327 |
+
3. If many frames show **"NO FACE"**, try brighter lighting or adjust your face angle.
|
| 328 |
+
4. Click **“Show Top 3 Surprise Moments”** after stopping the stream.
|
| 329 |
+
5. Download the generated PDF if needed.
|
| 330 |
+
---
|
| 331 |
+
"""
|
| 332 |
+
)
|
| 333 |
+
|
| 334 |
+
with gr.Row():
|
| 335 |
+
with gr.Column(scale=2):
|
| 336 |
+
|
| 337 |
+
webcam = gr.Image(
|
| 338 |
+
sources=["webcam"],
|
| 339 |
+
type="numpy",
|
| 340 |
+
label="Webcam Feed"
|
| 341 |
+
)
|
| 342 |
+
output_img = gr.Image(label="Detection Result")
|
| 343 |
+
|
| 344 |
+
with gr.Column(scale=1):
|
| 345 |
+
threshold = gr.Slider(
|
| 346 |
+
minimum=0.0, maximum=1.0, value=0.1,
|
| 347 |
+
step=0.01, label="Surprise Threshold"
|
| 348 |
+
)
|
| 349 |
+
|
| 350 |
+
gr.Markdown(
|
| 351 |
+
"""
|
| 352 |
+
### What is the Surprise Threshold?
|
| 353 |
+
|
| 354 |
+
- Lower threshold → detects smaller reactions
|
| 355 |
+
- Higher threshold → detects only strong surprise
|
| 356 |
+
- **Default = 0.1**
|
| 357 |
+
|
| 358 |
+
👉 Try making a surprised face to adjust sensitivity.
|
| 359 |
+
"""
|
| 360 |
+
)
|
| 361 |
+
|
| 362 |
+
output_label = gr.Label(label="Softmax Probabilities")
|
| 363 |
+
plot = gr.Plot(label="Emotion Probability (per frame)")
|
| 364 |
+
stats_md = gr.Markdown("### Session Stats\nWaiting for stream...")
|
| 365 |
+
|
| 366 |
+
webcam.stream(
|
| 367 |
+
fn=detect_surprise,
|
| 368 |
+
inputs=[webcam, threshold],
|
| 369 |
+
outputs=[output_img, output_label, plot, stats_md],
|
| 370 |
+
stream_every=0.1
|
| 371 |
+
)
|
| 372 |
+
|
| 373 |
+
gr.Markdown("---")
|
| 374 |
+
gr.Markdown("## 🔍 Summary & Report")
|
| 375 |
+
|
| 376 |
+
summarize_button = gr.Button("🎯 Show Top 3 Surprise Moments")
|
| 377 |
+
|
| 378 |
+
summary_text = gr.Textbox(
|
| 379 |
+
label="Top 3 Summary",
|
| 380 |
+
lines=6,
|
| 381 |
+
max_lines=10
|
| 382 |
+
)
|
| 383 |
+
|
| 384 |
+
with gr.Row():
|
| 385 |
+
img1 = gr.Image(label="Top 1")
|
| 386 |
+
img2 = gr.Image(label="Top 2")
|
| 387 |
+
img3 = gr.Image(label="Top 3")
|
| 388 |
+
|
| 389 |
+
timeline_plot = gr.Plot(label="Surprise Timeline")
|
| 390 |
+
pdf_file = gr.File(label="Download PDF Report")
|
| 391 |
+
|
| 392 |
+
summarize_button.click(
|
| 393 |
+
fn=summarize_results,
|
| 394 |
+
inputs=[],
|
| 395 |
+
outputs=[summary_text, img1, img2, img3, timeline_plot, pdf_file]
|
| 396 |
+
)
|
| 397 |
+
|
| 398 |
+
if __name__ == "__main__":
|
| 399 |
+
demo.launch()
|
|
|
|
|
|