Spaces:
Sleeping
Sleeping
File size: 19,177 Bytes
5412d82 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 | """
Production Gradio app β Gemma 4 E2B IT + Depth + YOLO.
Designed for real-world assistive use, not academic demonstration:
- Always runs Stage 3 (VLM + Depth + YOLO) β best spatial quality.
- Browser TTS via Web Speech API reads every new description aloud.
- Auto Mode fires inference automatically on scene change.
- Description history shows the last 3 outputs for quick review.
- Dark accessible theme with large text.
Requirements
------------
Run in the production conda environment (scene-prod), NOT the academic one.
See requirements_prod.txt for setup instructions.
Launch::
conda activate scene-prod
python -m src.ui.app_prod
Environment variables::
HF_TOKEN=hf_... # required for gated Gemma 4 model
PROD_SHARE=1 # set to expose a public Gradio tunnel
"""
from __future__ import annotations
import difflib
import textwrap
import time
from collections import deque
from typing import Any
import gradio as gr
import matplotlib
import numpy as np
from PIL import Image, ImageDraw, ImageFont
matplotlib.use("Agg")
import matplotlib.cm as cm # noqa: E402
from ..config import DEPTH_MAX_CM, DEPTH_MIN_CM
from ..depth_context import build_depth_context
from ..models.depth import DepthEstimator
from ..models.detector import ObjectDetector
from ..models.gemma4 import Gemma4VLM
from ..pipeline import Pipeline, _sync, _vram_mb
# ββ Constants ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
_FRAME_DIFF_THRESHOLD: float = 4.0 # grayscale MAD to trigger inference
_SIMILARITY_THRESHOLD: float = 0.80 # SequenceMatcher ratio to suppress TTS
_HISTORY_LEN: int = 3 # number of past descriptions to keep
_AUTO_INTERVAL_S: float = 6.0 # seconds between auto-describe ticks
_BOX_PALETTE: list[tuple[int, int, int]] = [
(255, 56, 56), # red
(56, 168, 255), # blue
(56, 255, 101), # green
(255, 178, 29), # orange
(180, 56, 255), # purple
(255, 56, 200), # pink
(29, 220, 220), # cyan
(255, 225, 56), # yellow
]
# Web Speech API JS snippet β injected via Gradio's js= parameter.
# Reads `text` aloud, cancelling any in-progress utterance first.
_TTS_JS = """
(text) => {
if (!text || text.trim() === "") return text;
window.speechSynthesis.cancel();
const u = new SpeechSynthesisUtterance(text);
u.rate = 0.92;
u.pitch = 1.0;
window.speechSynthesis.speak(u);
return text;
}
"""
# ββ Global pipeline (loaded once at startup) βββββββββββββββββββββββββββββββββ
_PIPELINE: Pipeline | None = None
def _get_pipeline() -> Pipeline:
global _PIPELINE
if _PIPELINE is None:
_PIPELINE = Pipeline(force_model="gemma4")
return _PIPELINE
# ββ AR overlay helpers ββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _draw_annotated(
frame_rgb: np.ndarray,
depth_np: np.ndarray | None,
boxes: np.ndarray,
classes: list[str],
confidences: list[float],
) -> np.ndarray:
"""Draw YOLO bounding boxes with class + distance labels."""
img = Image.fromarray(frame_rgb).convert("RGB")
draw = ImageDraw.Draw(img)
W, H = img.size
try:
font = ImageFont.truetype("arial.ttf", size=max(14, W // 50))
except (IOError, OSError):
font = ImageFont.load_default()
for i, (box, cls, conf) in enumerate(zip(boxes, classes, confidences)):
x1, y1, x2, y2 = (int(v) for v in box)
colour = _BOX_PALETTE[i % len(_BOX_PALETTE)]
for t in range(3):
draw.rectangle([x1 - t, y1 - t, x2 + t, y2 + t], outline=colour)
dist_str = ""
if depth_np is not None:
roi = depth_np[
max(0, y1):min(H - 1, y2) + 1,
max(0, x1):min(W - 1, x2) + 1,
]
if roi.size > 0:
med = float(np.median(roi))
dist_cm = DEPTH_MAX_CM - (med / 255.0) * (DEPTH_MAX_CM - DEPTH_MIN_CM)
dist_str = f" ~{dist_cm:.0f} cm"
label = f"{cls}{dist_str} {int(conf * 100)}%"
bbox = draw.textbbox((x1, y1), label, font=font)
pad = 3
draw.rectangle(
[bbox[0] - pad, bbox[1] - pad, bbox[2] + pad, bbox[3] + pad],
fill=colour,
)
draw.text((x1, y1), label, fill=(255, 255, 255), font=font)
return np.array(img)
def _depth_overlay(frame_rgb: np.ndarray, depth_np: np.ndarray, alpha: float = 0.50) -> np.ndarray:
"""Blend PLASMA colourmap of depth over the RGB frame."""
plasma = cm.get_cmap("plasma")
coloured = (plasma(depth_np / 255.0)[..., :3] * 255).astype(np.uint8)
return (
alpha * coloured.astype(np.float32)
+ (1.0 - alpha) * frame_rgb.astype(np.float32)
).astype(np.uint8)
def _placeholder(message: str, w: int = 640, h: int = 360) -> np.ndarray:
"""Dark-grey placeholder image with a centred message."""
base = np.full((h, w, 3), 30, dtype=np.uint8)
img = Image.fromarray(base).convert("RGB")
draw = ImageDraw.Draw(img)
try:
font = ImageFont.truetype("arial.ttf", size=max(16, w // 38))
except (IOError, OSError):
font = ImageFont.load_default()
for i, line in enumerate(textwrap.wrap(message, width=38)):
bbox = draw.textbbox((0, 0), line, font=font)
x = (w - (bbox[2] - bbox[0])) // 2
y = h // 2 - 20 + i * 28
draw.text((x + 1, y + 1), line, fill=(0, 0, 0), font=font)
draw.text((x, y), line, fill=(255, 220, 60), font=font)
return np.array(img)
# ββ Change detection ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _scene_changed(prev: np.ndarray | None, curr: np.ndarray) -> bool:
if prev is None:
return True
thumb = lambda a: np.array(Image.fromarray(a).resize((128, 72))).mean(axis=2)
return float(np.mean(np.abs(thumb(curr).astype(np.float32) - thumb(prev).astype(np.float32)))) >= _FRAME_DIFF_THRESHOLD
def _too_similar(a: str, b: str) -> bool:
if not a or not b:
return False
return difflib.SequenceMatcher(None, a.lower(), b.lower()).ratio() >= _SIMILARITY_THRESHOLD
# ββ Core inference βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _run_stage3(frame_rgb: np.ndarray) -> tuple[np.ndarray, np.ndarray, str, str, str]:
"""Run Stage 3 inference and return all UI outputs.
Returns:
annotated_frame, depth_map, description, depth_context, timing_text
"""
pipeline = _get_pipeline()
description, depth_context, timing = pipeline.run_stage3(frame_rgb)
# Reuse intermediates cached by run_stage3 β no second depth/YOLO pass
depth_np = pipeline.last_depth_np
boxes_np = pipeline.last_boxes if pipeline.last_boxes is not None else np.empty((0, 4), dtype=np.float32)
classes = pipeline.last_classes
confs = pipeline.last_confs
annotated = (
_draw_annotated(frame_rgb, depth_np, boxes_np, classes, confs)
if len(boxes_np) > 0
else frame_rgb.copy()
)
depth_vis = _depth_overlay(frame_rgb, depth_np) if depth_np is not None else frame_rgb.copy()
n_det = len(classes)
t = timing
timing_text = (
f"Total {t['total_s']*1000:.0f} ms | "
f"Depth {t.get('depth_s',0)*1000:.0f} ms | "
f"YOLO {t.get('yolo_s',0)*1000:.0f} ms ({n_det} obj) | "
f"VLM {t['vlm_s']*1000:.0f} ms | "
f"VRAM {t['vram_mb']:.0f} MB"
)
return annotated, depth_vis, description, depth_context, timing_text
# ββ Gradio callbacks ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def on_describe(
input_image: np.ndarray | None,
history: list[str],
) -> tuple:
"""Manual describe button callback."""
if input_image is None:
ph = _placeholder("Upload or capture an image first")
return ph, ph, "", "", "", history, ""
frame_rgb = input_image.astype(np.uint8)
annotated, depth_vis, description, depth_ctx, timing = _run_stage3(frame_rgb)
history = list(history)
if description:
history.insert(0, description)
history = history[:_HISTORY_LEN]
history_md = "\n\n---\n\n".join(
f"**[{i+1}]** {d}" for i, d in enumerate(history)
)
return annotated, depth_vis, description, depth_ctx, timing, history, history_md
def on_auto_tick(
live_frame: np.ndarray | None,
prev_description: str,
prev_frame_ref: np.ndarray | None,
history: list[str],
) -> tuple:
"""Auto Mode timer callback β runs only when the scene has changed."""
def _skip(status: str) -> tuple:
return (
gr.update(), gr.update(), # image panels
prev_description, # keep previous description
gr.update(), # depth context
f"βΈ {status}", # timing / status
prev_description, # state: prev_description
prev_frame_ref, # state: prev frame
history, # state: history
gr.update(), # history markdown
f"**{status}**", # status badge
)
if live_frame is None:
return _skip("Waiting for cameraβ¦")
frame_rgb = live_frame.astype(np.uint8)
if not _scene_changed(prev_frame_ref, frame_rgb):
return _skip("No change detected")
annotated, depth_vis, description, depth_ctx, timing = _run_stage3(frame_rgb)
if _too_similar(description, prev_description):
return (
annotated, depth_vis,
prev_description,
depth_ctx, timing,
prev_description, frame_rgb,
history, gr.update(),
"**Minor shift β description unchanged**",
)
new_history = list(history)
if description:
new_history.insert(0, description)
new_history = new_history[:_HISTORY_LEN]
history_md = "\n\n---\n\n".join(
f"**[{i+1}]** {d}" for i, d in enumerate(new_history)
)
return (
annotated, depth_vis,
description, depth_ctx, timing,
description, frame_rgb,
new_history, history_md,
"**Scene updated**",
)
def on_toggle_auto(is_active: bool) -> tuple:
"""Toggle the Auto Mode timer and swap visible webcam inputs."""
new_active = not is_active
label = "βΉ Stop Auto" if new_active else "βΆ Auto Mode"
status = "π΄ **Auto β describes every scene change**" if new_active else ""
return (
new_active,
gr.update(active=new_active),
gr.update(value=label),
status,
gr.update(visible=not new_active), # snapshot input
gr.update(visible=new_active), # streaming input
)
# ββ UI construction βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def build_ui() -> gr.Blocks:
"""Build and return the production Gradio Blocks interface."""
with gr.Blocks(title="Scene Describer β Gemma 4") as demo:
# ββ Header ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
gr.Markdown(
"""
# Scene Describer
### Spatial scene understanding for visually impaired users
Powered by **Gemma 4 E2B IT** Β· Depth Anything V2 Β· YOLOv8n
"""
)
# ββ Hidden state ββββββββββββββββββββββββββββββββββββββββββββββββββββββ
auto_active = gr.State(False)
prev_desc_st = gr.State("")
prev_frame_st = gr.State(None)
history_st = gr.State([])
timer = gr.Timer(value=_AUTO_INTERVAL_S, active=False)
with gr.Row():
# ββ Left: input controls ββββββββββββββββββββββββββββββββββββββββββ
with gr.Column(scale=1, min_width=300):
gr.Markdown("### Camera")
snapshot_in = gr.Image(
label="Upload or webcam snapshot",
sources=["webcam", "upload"],
type="numpy",
height=320,
)
streaming_in = gr.Image(
label="Live camera (Auto Mode)",
sources=["webcam"],
type="numpy",
streaming=True,
height=320,
visible=False,
)
with gr.Row():
describe_btn = gr.Button(
"Describe Scene",
variant="primary",
scale=2,
size="lg",
)
auto_btn = gr.Button(
"βΆ Auto Mode",
variant="secondary",
scale=1,
size="lg",
)
auto_status = gr.Markdown(
value="",
elem_classes=["status-badge"],
)
gr.Markdown(
"**Auto Mode** uses your live camera and describes the "
"scene whenever it changes. Each description is read "
"aloud automatically."
)
# ββ Right: outputs ββββββββββββββββββββββββββββββββββββββββββββββββ
with gr.Column(scale=2):
with gr.Row():
annotated_out = gr.Image(
label="Objects detected",
type="numpy",
height=250,
interactive=False,
)
depth_out = gr.Image(
label="Depth map",
type="numpy",
height=250,
interactive=False,
)
description_out = gr.Textbox(
label="Scene description",
lines=6,
interactive=False,
elem_classes=["description-box"],
placeholder="The description will appear here and be read aloudβ¦",
)
with gr.Accordion("Depth measurements (injected preamble)", open=False):
depth_ctx_out = gr.Textbox(
label="",
lines=6,
interactive=False,
show_label=False,
)
with gr.Accordion("Performance", open=False):
timing_out = gr.Textbox(
label="",
lines=2,
interactive=False,
show_label=False,
elem_classes=["timing-box"],
)
with gr.Accordion(f"Last {_HISTORY_LEN} descriptions", open=False):
history_md_out = gr.Markdown(
value="",
elem_classes=["history-box"],
)
# ββ Wiring ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# Manual describe button
describe_btn.click(
fn=on_describe,
inputs=[snapshot_in, history_st],
outputs=[
annotated_out, depth_out,
description_out, depth_ctx_out, timing_out,
history_st, history_md_out,
],
).then(
# After the Python callback, trigger TTS in the browser.
fn=None,
inputs=[description_out],
outputs=[description_out],
js=_TTS_JS,
)
# Auto Mode toggle
auto_btn.click(
fn=on_toggle_auto,
inputs=[auto_active],
outputs=[
auto_active, timer, auto_btn,
auto_status, snapshot_in, streaming_in,
],
)
# Timer tick
timer.tick(
fn=on_auto_tick,
inputs=[streaming_in, prev_desc_st, prev_frame_st, history_st],
outputs=[
annotated_out, depth_out,
description_out, depth_ctx_out, timing_out,
prev_desc_st, prev_frame_st,
history_st, history_md_out,
auto_status,
],
).then(
fn=None,
inputs=[description_out],
outputs=[description_out],
js=_TTS_JS,
)
return demo
# ββ Entry point βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def main() -> None:
"""Load the pipeline, build the UI, and launch."""
import os
# Eagerly load all models so the first request is fast.
print("Initialising pipeline (Gemma 4 + Depth + YOLO)β¦")
pipeline = _get_pipeline()
pipeline._get_depth()
pipeline._get_detector()
print("All models loaded. Starting UIβ¦")
demo = build_ui()
demo.launch(
server_name="127.0.0.1",
server_port=7862, # separate port from the academic app
share=bool(os.environ.get("PROD_SHARE", "")),
favicon_path=None,
theme=gr.themes.Base(
primary_hue=gr.themes.colors.violet,
secondary_hue=gr.themes.colors.slate,
neutral_hue=gr.themes.colors.slate,
font=gr.themes.GoogleFont("Inter"),
),
css="""
.description-box textarea { font-size: 1.15rem !important; line-height: 1.6 !important; }
.history-box { font-size: 0.95rem; }
.status-badge { font-size: 0.9rem; color: #aaa; }
.timing-box textarea { font-size: 0.8rem !important; font-family: monospace; }
""",
)
if __name__ == "__main__":
main()
|