File size: 16,136 Bytes
ebc1c85 | 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 | """Phase 1: Scene Understanding Module.
Combines:
- Metric depth estimation (Depth Anything V2)
- Room layout estimation (SpatialLM)
- Semantic segmentation (Mask2Former-style)
- Object detection & isolation (SAM)
"""
import os
from typing import Dict, List, Optional, Tuple, Union
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from PIL import Image
from transformers import (
pipeline,
AutoImageProcessor,
AutoModelForDepthEstimation,
AutoModelForSemanticSegmentation,
CLIPVisionModel,
CLIPImageProcessor,
)
class SceneUnderstandingModule(nn.Module):
"""Extract scene structure from a single interior image."""
def __init__(
self,
model_size: str = "L",
device: str = "cuda",
dtype: torch.dtype = torch.float16,
cache_dir: Optional[str] = None,
):
super().__init__()
self.model_size = model_size
self.device = device
self.dtype = dtype
self.cache_dir = cache_dir
# Load sub-models (lazy)
self._depth_model = None
self._depth_processor = None
self._segmentation_model = None
self._sam_model = None
self._room_classifier = None
@property
def depth_model(self):
if self._depth_model is None:
model_id = "depth-anything/Depth-Anything-V2-Metric-Indoor-Large-hf"
self._depth_processor = AutoImageProcessor.from_pretrained(
model_id, cache_dir=self.cache_dir
)
self._depth_model = AutoModelForDepthEstimation.from_pretrained(
model_id,
torch_dtype=self.dtype,
cache_dir=self.cache_dir,
).to(self.device)
self._depth_model.eval()
return self._depth_model, self._depth_processor
@property
def segmentation_model(self):
if self._segmentation_model is None:
# Using a generic indoor segmentation model
# In production, use fine-tuned Mask2Former or OneFormer
model_id = "facebook/mask2former-swin-large-coco-instance"
self._segmentation_model = AutoModelForSemanticSegmentation.from_pretrained(
model_id,
torch_dtype=self.dtype,
cache_dir=self.cache_dir,
).to(self.device)
self._segmentation_model.eval()
return self._segmentation_model
@property
def room_classifier(self):
if self._room_classifier is None:
self._room_classifier = CLIPVisionModel.from_pretrained(
"openai/clip-vit-large-patch14",
torch_dtype=self.dtype,
cache_dir=self.cache_dir,
).to(self.device)
self._room_classifier.eval()
return self._room_classifier
def forward(self, image: Image.Image) -> Dict:
"""
Process a single interior image and extract all scene understanding.
Returns:
Dictionary with keys:
- depth: metric depth map [H, W]
- normal: surface normal map [H, W, 3]
- room_layout: room layout structure
- semantic_segmentation: pixel-wise class labels [H, W]
- detected_objects: dict of per-object crops and masks
- room_type: str (e.g. "living_room")
- style: str (e.g. "modern")
"""
# Convert to tensor
img_np = np.array(image)
# === Metric Depth Estimation ===
depth_map = self.estimate_depth(image)
# === Surface Normal Estimation ===
normal_map = self.estimate_normals(depth_map)
# === Room Layout Estimation ===
room_layout = self.estimate_room_layout(depth_map, img_np)
# === Semantic Segmentation ===
semantic_seg = self.segment_image(image)
# === Object Detection & Isolation ===
detected_objects = self.detect_and_isolate_objects(
image, depth_map, semantic_seg
)
# === Room Type Classification ===
room_type = self.classify_room_type(image)
# === Style Classification ===
style = self.classify_style(image)
return {
"depth": depth_map,
"normal": normal_map,
"room_layout": room_layout,
"semantic_segmentation": semantic_seg,
"detected_objects": detected_objects,
"room_type": room_type,
"style": style,
}
def estimate_depth(self, image: Image.Image) -> np.ndarray:
"""Estimate metric depth using Depth Anything V2."""
model, processor = self.depth_model
inputs = processor(images=image, return_tensors="pt")
inputs = {k: v.to(self.device) for k, v in inputs.items()}
with torch.no_grad():
outputs = model(**inputs)
predicted_depth = outputs.predicted_depth
# Interpolate to original size
prediction = F.interpolate(
predicted_depth.unsqueeze(1),
size=image.size[::-1],
mode="bicubic",
align_corners=False,
)
depth = prediction.squeeze().cpu().numpy()
return depth
def estimate_normals(self, depth: np.ndarray) -> np.ndarray:
"""Compute surface normals from depth map."""
# Compute gradients
dz_dx = np.gradient(depth, axis=1)
dz_dy = np.gradient(depth, axis=0)
# Normal vector: [-dz/dx, -dz/dy, 1]
normals = np.stack([
-dz_dx,
-dz_dy,
np.ones_like(depth)
], axis=-1)
# Normalize
norm = np.linalg.norm(normals, axis=-1, keepdims=True)
normals = normals / (norm + 1e-8)
# Map to [0, 1] for visualization
normals_vis = (normals + 1) / 2
return normals
def estimate_room_layout(
self, depth: np.ndarray, img: np.ndarray
) -> Dict:
"""
Estimate room layout from depth map.
Detects:
- Floor plane (largest horizontal surface below camera)
- Ceiling plane (horizontal surface above, roughly at fixed height)
- Wall planes (vertical surfaces)
Returns Manhattan-world layout with planes.
"""
H, W = depth.shape
# Create point cloud from depth
# Assume standard camera intrinsics (can be refined later)
fx = fy = max(W, H) # approximate focal length
cx, cy = W / 2, H / 2
u, v = np.meshgrid(np.arange(W), np.arange(H))
z = depth
x = (u - cx) * z / fx
y = (v - cy) * z / fy
points = np.stack([x, y, z], axis=-1) # [H, W, 3]
# RANSAC plane detection
floor_plane = self._detect_floor_plane(points)
ceiling_plane = self._detect_ceiling_plane(points)
wall_planes = self._detect_wall_planes(points)
return {
"floor": floor_plane,
"ceiling": ceiling_plane,
"walls": wall_planes,
"point_cloud": points,
"dimensions": {
"width": float(np.max(x) - np.min(x)),
"depth": float(np.max(z) - np.min(z)),
"height": float(ceiling_plane["height"] - floor_plane["height"]),
}
}
def _detect_floor_plane(self, points: np.ndarray) -> Dict:
"""Detect floor plane (lowest y values, near-horizontal normal)."""
# Simple heuristic: lowest 20% of points
y_values = points[:, :, 1].flatten()
threshold = np.percentile(y_values, 20)
floor_mask = points[:, :, 1] < threshold
floor_points = points[floor_mask]
if len(floor_points) < 100:
# Fallback: assume floor at y=0
return {"normal": [0, 1, 0], "height": 0.0, "points": None}
# Fit plane
centroid = np.mean(floor_points, axis=0)
centered = floor_points - centroid
_, _, vh = np.linalg.svd(centered, full_matrices=False)
normal = vh[-1] # smallest singular value direction
# Ensure normal points up
if normal[1] < 0:
normal = -normal
return {
"normal": normal.tolist(),
"height": float(centroid[1]),
"centroid": centroid.tolist(),
"points": floor_points.shape[0],
}
def _detect_ceiling_plane(self, points: np.ndarray) -> Dict:
"""Detect ceiling plane (highest y values)."""
y_values = points[:, :, 1].flatten()
threshold = np.percentile(y_values, 90)
ceiling_mask = points[:, :, 1] > threshold
ceiling_points = points[ceiling_mask]
if len(ceiling_points) < 100:
# Fallback: typical room height ~2.7m
return {"normal": [0, -1, 0], "height": 2.7, "points": None}
centroid = np.mean(ceiling_points, axis=0)
return {
"normal": [0, -1, 0],
"height": float(centroid[1]),
"points": ceiling_points.shape[0],
}
def _detect_wall_planes(self, points: np.ndarray) -> List[Dict]:
"""Detect wall planes from remaining points."""
# Simplified: detect 4 walls for rectangular rooms
# In production, use proper RANSAC or SpatialLM
x = points[:, :, 0]
z = points[:, :, 2]
walls = []
# Left wall (minimum x)
x_min = np.percentile(x.flatten(), 5)
left_mask = np.abs(x - x_min) < 0.3
if np.sum(left_mask) > 100:
walls.append({
"normal": [1, 0, 0],
"position": float(x_min),
"direction": "left",
})
# Right wall (maximum x)
x_max = np.percentile(x.flatten(), 95)
right_mask = np.abs(x - x_max) < 0.3
if np.sum(right_mask) > 100:
walls.append({
"normal": [-1, 0, 0],
"position": float(x_max),
"direction": "right",
})
# Back wall (minimum z)
z_min = np.percentile(z.flatten(), 5)
back_mask = np.abs(z - z_min) < 0.3
if np.sum(back_mask) > 100:
walls.append({
"normal": [0, 0, 1],
"position": float(z_min),
"direction": "back",
})
# Front wall (maximum z)
z_max = np.percentile(z.flatten(), 95)
front_mask = np.abs(z - z_max) < 0.3
if np.sum(front_mask) > 100:
walls.append({
"normal": [0, 0, -1],
"position": float(z_max),
"direction": "front",
})
return walls
def segment_image(self, image: Image.Image) -> np.ndarray:
"""Run semantic segmentation to identify regions."""
# Placeholder: in production, use fine-tuned indoor segmentation
# For now, return a simple heuristic segmentation
img_np = np.array(image)
H, W = img_np.shape[:2]
# Heuristic: classify based on position and color
# Bottom 30% = floor, top 10% = ceiling, rest = walls + objects
seg = np.zeros((H, W), dtype=np.int32)
# Floor region
floor_threshold = int(H * 0.7)
seg[floor_threshold:] = 1 # floor
# Ceiling region
ceiling_threshold = int(H * 0.1)
seg[:ceiling_threshold] = 2 # ceiling
# Wall regions (sides)
wall_threshold = int(W * 0.15)
seg[ceiling_threshold:floor_threshold, :wall_threshold] = 3 # left wall
seg[ceiling_threshold:floor_threshold, -wall_threshold:] = 4 # right wall
return seg
def detect_and_isolate_objects(
self,
image: Image.Image,
depth: np.ndarray,
semantic_seg: np.ndarray,
) -> Dict:
"""
Detect and isolate furniture objects from the scene.
Returns dict mapping object_id -> {
crop: PIL Image,
mask: binary mask,
bbox: [x1, y1, x2, y2],
class_name: str,
depth_range: [min, max],
}
"""
# Placeholder: in production, use SAM + fine-tuned detector
# For now, return a simple grid-based detection
detected_objects = {}
img_np = np.array(image)
H, W = img_np.shape[:2]
# Divide room into zones and detect likely objects
# Floor zone: look for distinct objects above floor
floor_y = int(H * 0.65)
# Simple heuristic: detect high-gradient regions in floor area
from scipy.ndimage import label
floor_region = semantic_seg == 1
depth_floor = depth.copy()
depth_floor[~floor_region] = 0
# Find objects by depth discontinuities in floor region
depth_grad = np.abs(np.gradient(depth_floor)[0]) + \
np.abs(np.gradient(depth_floor)[1])
object_mask = depth_grad > np.percentile(depth_grad, 85)
labeled, num_features = label(object_mask)
for i in range(1, min(num_features + 1, 10)): # max 10 objects
obj_mask = labeled == i
ys, xs = np.where(obj_mask)
if len(xs) < 100: # skip tiny objects
continue
x1, y1 = int(xs.min()), int(ys.min())
x2, y2 = int(xs.max()), int(ys.max())
# Pad bbox
pad = 20
x1 = max(0, x1 - pad)
y1 = max(0, y1 - pad)
x2 = min(W, x2 + pad)
y2 = min(H, y2 + pad)
crop = image.crop((x1, y1, x2, y2))
mask_crop = obj_mask[y1:y2, x1:x2]
obj_depth = depth[obj_mask]
detected_objects[i - 1] = {
"crop": crop,
"mask": mask_crop,
"bbox": [x1, y1, x2, y2],
"class_name": "furniture", # would be classified in production
"depth_range": [float(obj_depth.min()), float(obj_depth.max())],
}
return detected_objects
def classify_room_type(self, image: Image.Image) -> str:
"""Classify room type from image."""
# Placeholder: use CLIP or fine-tuned classifier
# For now, return based on simple heuristics
img_np = np.array(image)
# Simple heuristic based on color distribution
# In production, use fine-tuned model
mean_color = img_np.mean(axis=(0, 1))
# Very simple heuristic (would be replaced with proper classifier)
if mean_color[2] > mean_color[0] + 20: # more blue = maybe kitchen/bathroom
return "kitchen"
elif mean_color[0] > mean_color[1] + 20: # more red = maybe bedroom
return "bedroom"
else:
return "living_room"
def classify_style(self, image: Image.Image) -> str:
"""Classify interior design style."""
# Placeholder: use fine-tuned style classifier
# Styles: modern, scandinavian, luxury, indian, commercial, minimalist
img_np = np.array(image)
mean_color = img_np.mean(axis=(0, 1))
std_color = img_np.std(axis=(0, 1))
# Simple heuristic (would be replaced with proper classifier)
if std_color.mean() < 30: # low color variation = minimalist/scandinavian
return "scandinavian"
elif mean_color.mean() > 180: # bright = modern
return "modern"
elif mean_color[0] < 80 and mean_color[1] < 80 and mean_color[2] < 80: # dark = luxury
return "luxury"
else:
return "modern"
|