MOD / app.py
Muhammad Usman
concurrent jobs serialized through one worker
9b19a6c
import asyncio
import base64
import io
import json
import os
import shutil
import time
try:
import tomllib
except ImportError:
try:
import tomli as tomllib
except ImportError:
try:
import tomlkit as tomllib
except ImportError:
raise ImportError(
"No TOML library found. Please run on Python 3.11+, or run 'pip install tomli' to support Python 3.10."
)
import uuid
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path
import cv2
import numpy as np
import torch
from fastapi import FastAPI, File, HTTPException, Response, UploadFile
from fastapi.middleware.cors import CORSMiddleware
from fastapi.staticfiles import StaticFiles
from PIL import Image, ImageOps
from transformers import (
AutoImageProcessor,
AutoModelForDepthEstimation,
Mask2FormerForUniversalSegmentation,
OneFormerForUniversalSegmentation,
OneFormerProcessor,
SegformerForSemanticSegmentation,
)
ADE20K_CLASSES = [
"wall", "building", "sky", "floor", "tree", "ceiling", "road", "bed",
"window", "grass", "cabinet", "sidewalk", "person", "ground", "door",
"table", "mountain", "plant", "curtain", "chair", "car", "water",
"painting", "sofa", "shelf", "house", "sea", "mirror", "rug", "field",
"armchair", "seat", "fence", "desk", "rock", "wardrobe", "lamp",
"bathtub", "railing", "cushion", "base", "box", "column", "signboard",
"chest of drawers", "counter", "sand", "sink", "skyscraper", "fireplace",
"refrigerator", "stairs", "runway", "bookcase", "blind", "coffee table",
"toilet", "flower", "book", "hill", "bench", "countertop", "stove",
"palm", "kitchen island", "computer", "swivel chair", "boat", "bar",
"arcade machine", "hovel", "bus", "towel", "light", "truck", "tower",
"chandelier", "awning", "streetlight", "booth", "television", "airplane",
"dirt track", "apparel", "pole", "land", "bannister", "escalator",
"ottoman", "bottle", "buffet", "poster", "stage", "van", "ship",
"fountain", "conveyer belt", "canopy", "washer", "plaything",
"swimming pool", "stool", "barrel", "basket", "waterfall", "tent",
"bag", "minibike", "cradle", "oven", "ball", "food", "step", "tank",
"trade name", "microwave", "pot", "animal", "bicycle", "lake",
"dishwasher", "screen", "blanket", "sculpture", "hood", "sconce",
"vase", "traffic light", "tray", "ashcan", "fan", "pier", "crt screen",
"plate", "monitor", "bulletin board", "shower", "radiator", "glass",
"clock", "flag",
]
def load_config() -> dict:
config_path = os.getenv("VISUALIZER_CONFIG")
if not config_path:
return {}
path = Path(config_path).expanduser()
if not path.is_absolute():
path = Path(__file__).resolve().parent / path
if not path.exists():
raise RuntimeError(f"VISUALIZER_CONFIG does not exist: {path}")
with path.open("rb") as config_file:
return tomllib.load(config_file)
CONFIG = load_config()
def config_value(env_name: str, section: str, key: str, default):
if env_name in os.environ:
return os.environ[env_name]
return CONFIG.get(section, {}).get(key, default)
SEGMENTATION_MODEL = str(
config_value("SEGMENTATION_MODEL", "models", "segmentation_model", "oneformer")
).lower()
ONEFORMER_MODEL_NAME = str(config_value(
"ONEFORMER_MODEL_NAME",
"models",
"oneformer_model_name",
"shi-labs/oneformer_ade20k_swin_large",
))
MASK2FORMER_MODEL_NAME = str(config_value(
"MASK2FORMER_MODEL_NAME",
"models",
"mask2former_model_name",
"facebook/mask2former-swin-small-ade-semantic",
))
SEGFORMER_MODEL_NAME = str(config_value(
"SEGFORMER_MODEL_NAME",
"models",
"segformer_model_name",
"nvidia/segformer-b2-finetuned-ade-512-512",
))
DEPTH_MODEL_NAME = str(config_value(
"DEPTH_MODEL_NAME",
"models",
"depth_model_name",
"Intel/dpt-large",
))
ENABLE_DEPTH_ESTIMATION = str(config_value(
"ENABLE_DEPTH_ESTIMATION",
"runtime",
"enable_depth_estimation",
"1",
)).lower() in {"1", "true", "yes", "on"}
INTRINSIC_MODEL_VERSION = str(config_value(
"INTRINSIC_MODEL_VERSION",
"models",
"intrinsic_model_version",
"v2",
))
ENABLE_INTRINSIC_SHADING = str(config_value(
"ENABLE_INTRINSIC_SHADING",
"runtime",
"enable_intrinsic_shading",
"0",
)).lower() in {"1", "true", "yes", "on"}
VISUALIZER_DATA_DIR = str(config_value(
"VISUALIZER_DATA_DIR",
"runtime",
"data_dir",
"data",
))
# Use the depth map (when available) to estimate the floor plane / homography
# instead of the geometric edge-fit. The edge-fit only sees perspective when the
# floor's side boundaries converge inside the frame, so it badly under-estimates
# perspective for floors that fill the frame. The depth-based fit recovers true
# foreshortening from the depth gradient. Edge-fit remains the fallback.
FLOOR_PLANE_FROM_DEPTH = str(config_value(
"FLOOR_PLANE_FROM_DEPTH",
"runtime",
"floor_plane_from_depth",
"1",
)).lower() in {"1", "true", "yes", "on"}
# Where to place the floor's horizon relative to the visible depth range, as a
# fraction of that range beyond the farthest visible floor pixel. Smaller -> the
# horizon sits closer to the visible floor -> stronger foreshortening. Larger ->
# milder perspective. This is the main knob for tuning perspective strength.
FLOOR_HORIZON_GAMMA = float(config_value(
"FLOOR_HORIZON_GAMMA",
"runtime",
"floor_horizon_gamma",
"0.25",
))
# Assumed pinhole focal length as a multiple of the image's long side. Only
# affects the cross-floor (X) scale / tile aspect, not foreshortening strength.
FLOOR_FOCAL_FACTOR = float(config_value(
"FLOOR_FOCAL_FACTOR",
"runtime",
"floor_focal_factor",
"1.0",
))
# Number of concurrent inference jobs. The segmentation/depth models are shared,
# global, and not thread-safe, and a single GPU/CPU can only run one at a time,
# so this defaults to 1 (fully serialized). Raise only with per-worker models.
INFERENCE_WORKERS = max(1, int(config_value(
"INFERENCE_WORKERS",
"runtime",
"inference_workers",
"1",
)))
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
seg_processor = None
seg_model = None
segmentation_backend = "segformer"
depth_processor = None
depth_model = None
intrinsic_models = None
def hf_offline() -> bool:
return os.getenv("HF_HUB_OFFLINE") == "1" or os.getenv("TRANSFORMERS_OFFLINE") == "1"
def _load_segmentation_model():
global seg_processor, seg_model, segmentation_backend
if SEGMENTATION_MODEL == "oneformer":
try:
print(f"Loading OneFormer: {ONEFORMER_MODEL_NAME} ...", flush=True)
start_time = time.perf_counter()
seg_processor = OneFormerProcessor.from_pretrained(
ONEFORMER_MODEL_NAME,
local_files_only=hf_offline(),
)
seg_model = OneFormerForUniversalSegmentation.from_pretrained(
ONEFORMER_MODEL_NAME,
local_files_only=hf_offline(),
).to(device)
seg_model.eval()
segmentation_backend = "oneformer"
print(f"OneFormer loaded in {time.perf_counter() - start_time:.4f}s.", flush=True)
return
except Exception as exc:
print(f"OneFormer failed ({exc}), falling back to Mask2Former.", flush=True)
if SEGMENTATION_MODEL in {"oneformer", "mask2former"}:
try:
print(f"Loading Mask2Former: {MASK2FORMER_MODEL_NAME} ...", flush=True)
start_time = time.perf_counter()
seg_processor = AutoImageProcessor.from_pretrained(
MASK2FORMER_MODEL_NAME,
local_files_only=hf_offline(),
)
seg_model = Mask2FormerForUniversalSegmentation.from_pretrained(
MASK2FORMER_MODEL_NAME,
local_files_only=hf_offline(),
).to(device)
seg_model.eval()
segmentation_backend = "mask2former"
print(f"Mask2Former loaded in {time.perf_counter() - start_time:.4f}s.", flush=True)
return
except Exception as exc:
print(f"Mask2Former failed ({exc}), falling back to SegFormer.", flush=True)
print(f"Loading SegFormer: {SEGFORMER_MODEL_NAME} ...", flush=True)
start_time = time.perf_counter()
seg_processor = AutoImageProcessor.from_pretrained(
SEGFORMER_MODEL_NAME,
local_files_only=hf_offline(),
)
seg_model = SegformerForSemanticSegmentation.from_pretrained(
SEGFORMER_MODEL_NAME,
local_files_only=hf_offline(),
).to(device)
seg_model.eval()
segmentation_backend = "segformer"
print(f"SegFormer loaded in {time.perf_counter() - start_time:.4f}s.", flush=True)
_load_segmentation_model()
def _load_intrinsic_model():
global intrinsic_models
if ENABLE_INTRINSIC_SHADING and intrinsic_models is None:
try:
print(f"Loading Intrinsic Image Decomposition model: {INTRINSIC_MODEL_VERSION} ...", flush=True)
start_time = time.perf_counter()
from intrinsic.pipeline import load_models
intrinsic_models = load_models(INTRINSIC_MODEL_VERSION, device=str(device))
print(f"Intrinsic model loaded in {time.perf_counter() - start_time:.4f}s.", flush=True)
except Exception as exc:
print(f"Intrinsic model failed to load ({exc}). Falling back to luminance shading.", flush=True)
_load_intrinsic_model()
def _load_depth_model():
global depth_processor, depth_model
if ENABLE_DEPTH_ESTIMATION and (depth_processor is None or depth_model is None):
try:
model_name = DEPTH_MODEL_NAME
print(f"Loading depth model: {model_name} ...", flush=True)
start_time = time.perf_counter()
depth_processor = AutoImageProcessor.from_pretrained(
model_name,
local_files_only=hf_offline(),
)
depth_model = AutoModelForDepthEstimation.from_pretrained(
model_name,
local_files_only=hf_offline(),
).to(device)
depth_model.eval()
print(f"Depth model loaded in {time.perf_counter() - start_time:.4f}s.", flush=True)
except Exception as exc:
print(f"Depth model failed to load ({exc}).", flush=True)
_load_depth_model()
app = FastAPI()
app.add_middleware(
CORSMiddleware,
allow_origins=["*"],
allow_methods=["*"],
allow_headers=["*"],
)
DATA_DIR = Path(VISUALIZER_DATA_DIR).resolve()
UPLOAD_DIR = DATA_DIR / "uploads"
JOB_DIR = DATA_DIR / "jobs"
UPLOAD_DIR.mkdir(parents=True, exist_ok=True)
JOB_DIR.mkdir(parents=True, exist_ok=True)
app.mount("/uploads", StaticFiles(directory=UPLOAD_DIR), name="uploads")
# All heavy model inference runs here, off the event loop and serialized to
# INFERENCE_WORKERS (default 1) so the shared, non-thread-safe models are never
# run concurrently. Both /segment and the background /viz2d/convert jobs submit
# to this single pool, bounding total in-flight inference regardless of traffic.
INFERENCE_POOL = ThreadPoolExecutor(
max_workers=INFERENCE_WORKERS,
thread_name_prefix="inference",
)
PRIMARY_FLOOR_CLASSES = {"floor"}
FLOOR_SURFACE_CLASSES = {
"floor", "road", "sidewalk", "ground", "field", "grass", "sand",
"runway", "dirt track", "land", "stairs", "step",
}
REJECT_SURFACE_CLASSES = {"wall", "ceiling", "building", "sky", "window"}
OCCLUDER_CLASSES = {
"bed", "cabinet", "person", "door", "table", "plant", "curtain", "chair",
"car", "painting", "sofa", "shelf", "mirror", "rug", "armchair", "seat", "desk",
"wardrobe", "lamp", "bathtub", "railing", "cushion", "base", "box",
"column", "chest of drawers", "counter", "sink", "fireplace",
"refrigerator", "bookcase", "blind", "coffee table", "toilet", "bench",
"countertop", "stove", "kitchen island", "computer", "swivel chair",
"bar", "ottoman", "bottle", "buffet", "poster", "towel", "television",
"washer", "plaything", "stool", "basket", "bag", "cradle", "oven",
"ball", "food", "microwave", "pot", "dishwasher", "blanket", "sculpture",
"vase", "tray", "fan", "plate", "monitor", "shower", "radiator", "clock",
}
def class_name_for_id(class_id: int) -> str:
return ADE20K_CLASSES[class_id] if class_id < len(ADE20K_CLASSES) else f"class_{class_id}"
def class_ids(names: set[str]) -> list[int]:
return [idx for idx, name in enumerate(ADE20K_CLASSES) if name in names]
def estimate_depth(img: Image.Image, width: int, height: int, task_id: str = "segment"):
global depth_processor, depth_model
if not ENABLE_DEPTH_ESTIMATION:
return None
model_name = DEPTH_MODEL_NAME
print(f"[{task_id}] Starting depth estimation...", flush=True)
start_time = time.perf_counter()
try:
if depth_processor is None or depth_model is None:
print(f"[{task_id}] Loading depth model: {model_name} ...", flush=True)
start_load = time.perf_counter()
depth_processor = AutoImageProcessor.from_pretrained(
model_name,
local_files_only=hf_offline(),
)
depth_model = AutoModelForDepthEstimation.from_pretrained(
model_name,
local_files_only=hf_offline(),
).to(device)
depth_model.eval()
print(f"[{task_id}] Depth model loaded in {time.perf_counter() - start_load:.4f}s.", flush=True)
inputs = depth_processor(images=img, return_tensors="pt").to(device)
with torch.no_grad():
outputs = depth_model(**inputs)
depth = torch.nn.functional.interpolate(
outputs.predicted_depth.unsqueeze(1),
size=(height, width),
mode="bicubic",
align_corners=False,
).squeeze().cpu().numpy()
depth = cv2.GaussianBlur(depth.astype(np.float32), (0, 0), sigmaX=3)
depth_min, depth_max = float(np.min(depth)), float(np.max(depth))
duration = time.perf_counter() - start_time
print(f"[{task_id}] Depth estimation completed in {duration:.4f}s", flush=True)
if depth_max - depth_min < 1e-6:
return None
return (depth - depth_min) / (depth_max - depth_min)
except Exception as exc:
print(f"[{task_id}] Depth estimation skipped ({exc}).", flush=True)
return None
def build_shade_map(img_np: np.ndarray, surface_mask: np.ndarray, task_id: str = "segment") -> np.ndarray | None:
if not surface_mask.any():
return None
print(f"[{task_id}] Starting shade map build...", flush=True)
start_time = time.perf_counter()
mask = surface_mask.astype(np.uint8)
luminance = (
img_np[:, :, 0].astype(np.float32) * 0.299
+ img_np[:, :, 1].astype(np.float32) * 0.587
+ img_np[:, :, 2].astype(np.float32) * 0.114
)
h, w = mask.shape[:2]
floor_values = luminance[mask > 0]
if floor_values.size < max(256, int(h * w * 0.002)):
return None
median_lum = float(np.median(floor_values))
if median_lum < 1e-3:
return None
filled = luminance.copy()
filled[mask == 0] = median_lum
missing = (mask == 0).astype(np.uint8) * 255
try:
filled = cv2.inpaint(
np.clip(filled, 0, 255).astype(np.uint8),
missing,
max(3, min(h, w) // 160),
cv2.INPAINT_TELEA,
).astype(np.float32)
except cv2.error:
pass
sigma = max(8.0, min(h, w) / 28.0)
smooth = cv2.GaussianBlur(filled, (0, 0), sigmaX=sigma, sigmaY=sigma)
shade = np.clip(smooth / median_lum, 0.55, 1.35)
shade[mask == 0] = 1.0
result = np.round((shade - 0.55) * (255.0 / 0.80)).clip(0, 255).astype(np.uint8)
duration = time.perf_counter() - start_time
print(f"[{task_id}] Shade map built in {duration:.4f}s", flush=True)
return result
def build_intrinsic_shade_map(img_np: np.ndarray, surface_mask: np.ndarray, task_id: str = "segment") -> np.ndarray | None:
if not surface_mask.any() or intrinsic_models is None:
return None
print(f"[{task_id}] Starting intrinsic shade map build...", flush=True)
start_time = time.perf_counter()
try:
# Convert image to float32 range [0.0, 1.0] as expected by compphoto/Intrinsic
img_float = img_np.astype(np.float32) / 255.0
# Run pipeline
from intrinsic.pipeline import run_pipeline
# Use CPU/CUDA device string
results = run_pipeline(intrinsic_models, img_float, stage=1, device=str(device))
# Extract shading map
shading = None
if "gry_shd" in results:
shading = results["gry_shd"]
elif "dif_shd" in results:
# If 3-channel diffuse shading, convert to grayscale luminance
dif = results["dif_shd"]
shading = dif[:, :, 0] * 0.299 + dif[:, :, 1] * 0.587 + dif[:, :, 2] * 0.114
else:
# Check other fallback keys matching "shd" or "shading"
for k in results.keys():
if "shd" in k or "shading" in k:
shading = results[k]
if len(shading.shape) == 3:
shading = shading[:, :, 0] * 0.299 + shading[:, :, 1] * 0.587 + shading[:, :, 2] * 0.114
break
if shading is None:
return None
# Resize shading to match original image size if there's any mismatch
h, w = surface_mask.shape[:2]
if shading.shape[:2] != (h, w):
shading = cv2.resize(shading, (w, h), interpolation=cv2.INTER_LINEAR)
# Smooth the shading map to eliminate high-frequency grout lines/patterns
# while preserving broad ambient shadows (e.g. plant shadows).
sigma = max(3.0, min(h, w) / 80.0)
shading = cv2.GaussianBlur(shading.astype(np.float32), (0, 0), sigmaX=sigma, sigmaY=sigma)
floor_vals = shading[surface_mask > 0]
if floor_vals.size == 0:
return None
median_val = float(np.median(floor_vals))
if median_val < 1e-3:
return None
# Divide by median to get relative shading multiplier (1.0 is neutral)
relative_shading = shading / median_val
# Clip relative shading to [0.55, 1.35]
relative_shading = np.clip(relative_shading, 0.55, 1.35)
# Non-floor pixels are neutral (1.0)
relative_shading[surface_mask == 0] = 1.0
# Encode to [0, 255] byte range matching the frontend
result = np.round((relative_shading - 0.55) * (255.0 / 0.80)).clip(0, 255).astype(np.uint8)
duration = time.perf_counter() - start_time
print(f"[{task_id}] Intrinsic shade map built in {duration:.4f}s", flush=True)
return result
except Exception as exc:
print(f"[{task_id}] Intrinsic shading decomposition failed: {exc}. Falling back to default luminance shading.", flush=True)
return None
def clean_floor_mask(mask: np.ndarray) -> np.ndarray:
if mask.dtype != np.uint8:
mask = mask.astype(np.uint8)
h, w = mask.shape[:2]
min_side = max(3, min(h, w))
close_size = max(5, int(round(min_side * 0.018))) | 1
open_size = max(3, int(round(min_side * 0.006))) | 1
closed = cv2.morphologyEx(
mask,
cv2.MORPH_CLOSE,
cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (close_size, close_size)),
)
cleaned = cv2.morphologyEx(
closed,
cv2.MORPH_OPEN,
cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (open_size, open_size)),
)
count, labels, stats, _ = cv2.connectedComponentsWithStats(cleaned, connectivity=8)
if count <= 1:
return cleaned
gravity_threshold = int(h * 0.60)
min_area = max(1000, int(h * w * 0.01))
result = np.zeros_like(cleaned)
for component_id in range(1, count):
area = stats[component_id, cv2.CC_STAT_AREA]
if area < min_area:
continue
comp_bottom = stats[component_id, cv2.CC_STAT_TOP] + stats[component_id, cv2.CC_STAT_HEIGHT]
if comp_bottom <= gravity_threshold:
continue
result[labels == component_id] = 1
if result.any():
return result
largest = 1 + int(np.argmax(stats[1:, cv2.CC_STAT_AREA]))
return (labels == largest).astype(np.uint8)
def wall_subtract(mask: np.ndarray, seg_map: np.ndarray, dilation: int = 1) -> np.ndarray:
reject_raw = np.isin(seg_map, class_ids(REJECT_SURFACE_CLASSES)).astype(np.uint8)
if dilation > 0:
kern = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
reject_raw = cv2.dilate(reject_raw, kern, iterations=dilation)
result = mask.copy()
result[reject_raw > 0] = 0
return result
def fit_floor_edges(mask: np.ndarray):
h, w = mask.shape[:2]
row_ys, lefts, rights = [], [], []
step = max(1, h // 260)
for y in range(0, h, step):
row_xs = np.where(mask[y] > 0)[0]
if len(row_xs) < max(8, w * 0.01):
continue
row_ys.append(float(y))
lefts.append(float(np.percentile(row_xs, 3)))
rights.append(float(np.percentile(row_xs, 97)))
if len(row_ys) < 8:
return None
row_ys_np = np.asarray(row_ys, dtype=np.float32)
return np.polyfit(row_ys_np, np.asarray(lefts, dtype=np.float32), 1), np.polyfit(
row_ys_np,
np.asarray(rights, dtype=np.float32),
1,
)
def detect_vanishing_point(img_np: np.ndarray, floor_mask: np.ndarray):
gray = cv2.cvtColor(img_np, cv2.COLOR_RGB2GRAY)
gray = cv2.GaussianBlur(gray, (5, 5), 0)
edges = cv2.Canny(gray, 60, 160)
edges[floor_mask == 0] = 0
lines = cv2.HoughLinesP(
edges,
rho=1,
theta=np.pi / 180,
threshold=60,
minLineLength=max(40, min(img_np.shape[:2]) // 16),
maxLineGap=24,
)
if lines is None:
return None
h, w = img_np.shape[:2]
candidates = []
for line in lines[:, 0, :]:
x1, y1, x2, y2 = [float(v) for v in line]
dx, dy = x2 - x1, y2 - y1
length = float(np.hypot(dx, dy))
if length < 40 or abs(dx) < 1:
continue
slope = dy / dx
if abs(slope) >= 0.18:
candidates.append((x1, y1, x2, y2, slope, length))
intersections = []
for i, (x1, y1, _, _, s1, l1) in enumerate(candidates):
a1 = y1 - s1 * x1
for x3, y3, _, _, s2, l2 in candidates[i + 1:]:
if s1 * s2 > 0 or abs(s1 - s2) < 0.12:
continue
a2 = y3 - s2 * x3
x = (a2 - a1) / (s1 - s2)
y = s1 * x + a1
if -w * 0.5 <= x <= w * 1.5 and -h <= y <= h * 0.95:
intersections.append((x, y, min(l1, l2)))
if not intersections:
return None
pts = np.asarray([[p[0], p[1]] for p in intersections], dtype=np.float32)
weights = np.asarray([p[2] for p in intersections], dtype=np.float32)
center = np.average(pts, axis=0, weights=weights)
dist = np.linalg.norm(pts - center, axis=1)
keep = dist <= np.percentile(dist, 70)
if keep.sum() >= 3:
center = np.average(pts[keep], axis=0, weights=weights[keep])
return {"x": float(center[0]), "y": float(center[1])}
def _fit_disparity_plane(xs: np.ndarray, ys: np.ndarray, d: np.ndarray):
"""Least-squares fit of d ~= a*x + b*y + c with one robust reweighting pass.
For a planar floor the (inverse-)depth a DPT-style model produces is an
affine function of image coordinates, so this recovers the plane directly.
"""
A = np.stack([xs, ys, np.ones_like(xs)], axis=1)
coeffs, *_ = np.linalg.lstsq(A, d, rcond=None)
resid = d - A @ coeffs
scale = float(np.median(np.abs(resid))) + 1e-9
keep = np.abs(resid) < 3.0 * scale
if int(keep.sum()) > A.shape[1] * 8:
coeffs, *_ = np.linalg.lstsq(A[keep], d[keep], rcond=None)
return float(coeffs[0]), float(coeffs[1]), float(coeffs[2])
def estimate_floor_plane_from_depth(mask: np.ndarray, depth: np.ndarray, task_id: str = "segment"):
"""Estimate the floor homography from the depth map.
A planar floor's inverse-depth d is affine in pixel coords, d = a*u + b*v + c,
so the map image-pixel -> world floor coords (X, Z) is the homography
H = [[1/f, 0, -cx/f],
[0, 0, 1 ],
[a, b, c-b ]] (b = horizon offset beta)
where Z = 1/(a*u+b*v+(c-beta)) is world depth and X = (u-cx)/(f*Z) is the
world cross-floor coordinate. Tiling uniformly in (X, Z) yields correct
perspective foreshortening. Returns (homography_flat, plane) in the same
format as estimate_floor_plane, or (None, None) if depth is uninformative.
"""
print(f"[{task_id}] Starting depth-based floor plane estimation...", flush=True)
start_time = time.perf_counter()
h, w = mask.shape[:2]
ys, xs = np.where(mask > 0)
if len(xs) < 1000:
return None, None
xs_f = xs.astype(np.float64)
ys_f = ys.astype(np.float64)
d_all = depth[ys, xs].astype(np.float64)
# Subsample for the fit to keep it fast on full floors.
if len(xs_f) > 150000:
idx = np.linspace(0, len(xs_f) - 1, 150000).astype(np.int64)
a, b, c = _fit_disparity_plane(xs_f[idx], ys_f[idx], d_all[idx])
else:
a, b, c = _fit_disparity_plane(xs_f, ys_f, d_all)
if np.hypot(a, b) < 1e-9:
# Disparity is ~constant across the floor: no usable perspective signal.
print(f"[{task_id}] Depth plane fit degenerate (flat disparity), falling back.", flush=True)
return None, None
# Predicted disparity over the floor; place the horizon GAMMA*range beyond the
# farthest visible pixel so the denominator stays strictly positive everywhere.
d_pred = a * xs_f + b * ys_f + c
d_min = float(d_pred.min())
d_range = float(np.percentile(d_pred, 99.5) - np.percentile(d_pred, 0.5))
if d_range < 1e-9:
return None, None
beta = d_min - max(FLOOR_HORIZON_GAMMA, 1e-3) * d_range
cc = c - beta
cx, cy = w * 0.5, h * 0.5
f = max(FLOOR_FOCAL_FACTOR * max(w, h), 1.0)
H = np.array([
[1.0 / f, 0.0, -cx / f],
[0.0, 0.0, 1.0],
[a, b, cc],
], dtype=np.float64)
denom = a * xs_f + b * ys_f + cc
if float(denom.min()) <= 1e-6:
# Numerical safety: keep every floor pixel strictly in front of the horizon.
return None, None
fx = (xs_f / f - cx / f) / denom
fy = 1.0 / denom
x1, x2 = float(np.percentile(fx, 1)), float(np.percentile(fx, 99))
y1, y2 = float(np.percentile(fy, 1)), float(np.percentile(fy, 99))
width, height = x2 - x1, y2 - y1
if width < 1e-9 or height < 1e-9:
return None, None
# The rectified coords above are in world/disparity units (tens), but the
# frontend sizes tiles from plane.width/height assuming image-pixel scale
# (like the legacy edge-fit). Without this the tiles come out enormous.
# Rescale the homography output (rows 0/1) and the plane rect to ~pixel scale.
target = float(max(w, h))
k = target / max(width, height)
H[0, :] *= k
H[1, :] *= k
x1, x2, y1, y2 = x1 * k, x2 * k, y1 * k, y2 * k
width, height = width * k, height * k
# Image-space quad = rectified bbox corners mapped back through H^-1.
try:
h_inv = np.linalg.inv(H)
except np.linalg.LinAlgError:
return None, None
rect = np.array([
[x1, y2, 1.0],
[x2, y2, 1.0],
[x2, y1, 1.0],
[x1, y1, 1.0],
], dtype=np.float64).T
quad_h = h_inv @ rect
quad = (quad_h[:2] / quad_h[2]).T
quad[:, 0] = np.clip(quad[:, 0], 0, w - 1)
quad[:, 1] = np.clip(quad[:, 1], 0, h - 1)
# Vanishing point: image location of the horizon along the recede direction
# (denominator -> 0), for debugging / downstream consumers.
vanishing_point = None
norm2 = a * a + b * b
if norm2 > 1e-12:
vp_x = cx - a * cc / norm2
vp_y = cy - b * cc / norm2
vanishing_point = {"x": float(vp_x), "y": float(vp_y)}
homography = H.flatten().tolist()
duration = time.perf_counter() - start_time
print(f"[{task_id}] Depth-based floor plane estimation completed in {duration:.4f}s "
f"(grad=({a:.3e},{b:.3e}), beta={beta:.4f}).", flush=True)
return homography, {
"x": x1,
"y": y1,
"width": width,
"height": height,
"quad": quad.flatten().tolist(),
"vanishingPoint": vanishing_point,
}
def estimate_floor_plane(mask: np.ndarray, img_np: np.ndarray, task_id: str = "segment"):
print(f"[{task_id}] Starting floor plane estimation...", flush=True)
start_time = time.perf_counter()
ys, xs = np.where(mask > 0)
if len(xs) < 1000:
return None, None
xs_f, ys_f = xs.astype(np.float32), ys.astype(np.float32)
x1, x2 = float(np.percentile(xs_f, 1)), float(np.percentile(xs_f, 99))
y1, y2 = float(np.percentile(ys_f, 1)), float(np.percentile(ys_f, 99))
width, height = x2 - x1, y2 - y1
if width < 20 or height < 20:
return None, None
top_y = float(np.percentile(ys_f, 8))
bottom_y = float(np.percentile(ys_f, 97))
edge_fits = fit_floor_edges(mask)
if edge_fits is None:
return None, None
left_fit, right_fit = edge_fits
top_left = float(np.polyval(left_fit, top_y))
top_right = float(np.polyval(right_fit, top_y))
bottom_left = float(np.polyval(left_fit, bottom_y))
bottom_right = float(np.polyval(right_fit, bottom_y))
lower_xs = xs_f[ys_f >= np.percentile(ys_f, 80)]
bottom_left = min(bottom_left, float(np.percentile(lower_xs, 4)))
bottom_right = max(bottom_right, float(np.percentile(lower_xs, 96)))
min_top_width = max(24.0, width * 0.18)
top_center = (top_left + top_right) * 0.5
if top_right - top_left < min_top_width:
top_left = top_center - min_top_width * 0.5
top_right = top_center + min_top_width * 0.5
min_bottom_width = max(min_top_width * 1.25, width * 0.45)
bottom_center = (bottom_left + bottom_right) * 0.5
if bottom_right - bottom_left < min_bottom_width:
bottom_left = bottom_center - min_bottom_width * 0.5
bottom_right = bottom_center + min_bottom_width * 0.5
h, w = mask.shape[:2]
src = np.float32([
[np.clip(bottom_left, 0, w - 1), np.clip(bottom_y, 0, h - 1)],
[np.clip(bottom_right, 0, w - 1), np.clip(bottom_y, 0, h - 1)],
[np.clip(top_right, 0, w - 1), np.clip(top_y, 0, h - 1)],
[np.clip(top_left, 0, w - 1), np.clip(top_y, 0, h - 1)],
])
vanishing_point = detect_vanishing_point(img_np, mask)
if vanishing_point is not None and vanishing_point["y"] < bottom_y:
vp_x = float(np.clip(vanishing_point["x"], -w * 0.25, w * 1.25))
top_width = max(src[2][0] - src[3][0], width * 0.16)
horizon_gap = max(bottom_y - top_y, 1.0)
convergence = np.clip((top_y - vanishing_point["y"]) / horizon_gap, 0.12, 0.75)
top_center = top_center * (1 - convergence * 0.35) + vp_x * (convergence * 0.35)
src[3][0] = np.clip(top_center - top_width * 0.5, 0, w - 1)
src[2][0] = np.clip(top_center + top_width * 0.5, 0, w - 1)
if cv2.contourArea(src) < 100:
return None, None
dst = np.float32([[x1, y2], [x2, y2], [x2, y1], [x1, y1]])
homography = cv2.getPerspectiveTransform(src, dst).flatten().tolist()
duration = time.perf_counter() - start_time
print(f"[{task_id}] Floor plane estimation completed in {duration:.4f}s", flush=True)
return homography, {
"x": x1,
"y": y1,
"width": width,
"height": height,
"quad": src.flatten().tolist(),
"vanishingPoint": vanishing_point,
}
def build_floor_surface_mask(
floor_mask: np.ndarray,
seg_map: np.ndarray,
quad: np.ndarray | None,
depth: np.ndarray | None,
task_id: str = "segment",
):
print(f"[{task_id}] Starting floor surface mask build...", flush=True)
start_time = time.perf_counter()
h, w = floor_mask.shape[:2]
kern_size = max(5, min(h, w) // 160) | 1
kern = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (kern_size, kern_size))
occluder_mask = np.isin(seg_map, class_ids(OCCLUDER_CLASSES)).astype(np.uint8)
occ_dilated = cv2.dilate(occluder_mask, kern, iterations=2)
reject_mask = np.isin(seg_map, class_ids(REJECT_SURFACE_CLASSES)).astype(np.uint8)
reject_dilated = cv2.dilate(reject_mask, kern, iterations=2)
surface = floor_mask.copy()
surface[reject_dilated > 0] = 0
if not surface.any():
surface = floor_mask.copy()
contours, _ = cv2.findContours(surface, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
if contours:
filled = np.zeros((h, w), dtype=np.uint8)
cv2.drawContours(filled, contours, -1, 1, cv2.FILLED)
filled[reject_dilated > 0] = 0
surface = filled
if quad is not None and surface.any():
plane_mask = np.zeros((h, w), dtype=np.uint8)
cv2.fillConvexPoly(plane_mask, np.round(quad).astype(np.int32), 1)
plane_mask[reject_dilated > 0] = 0
near_floor = cv2.dilate(surface, kern, iterations=6)
surface = cv2.bitwise_or(surface, cv2.bitwise_and(plane_mask, near_floor))
surface[occ_dilated > 0] = 0
if depth is not None and floor_mask.any():
floor_depth = depth[floor_mask > 0]
lo, hi = float(np.percentile(floor_depth, 2)), float(np.percentile(floor_depth, 98))
margin = max(0.08, (hi - lo) * 0.35)
depth_keep = (depth >= lo - margin) & (depth <= hi + margin)
surface = (surface & depth_keep.astype(np.uint8)).astype(np.uint8)
surface[floor_mask > 0] = np.maximum(surface[floor_mask > 0], 1)
surface[occ_dilated > 0] = 0
surface[reject_dilated > 0] = 0
surface = clean_floor_mask(surface)
surface[occ_dilated > 0] = 0
surface[reject_dilated > 0] = 0
boundary_kern = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3))
surface = cv2.dilate(surface, boundary_kern, iterations=1)
surface[occ_dilated > 0] = 0
surface[reject_dilated > 0] = 0
duration = time.perf_counter() - start_time
print(f"[{task_id}] Floor surface mask built in {duration:.4f}s", flush=True)
return surface
def run_segmentation(img: Image.Image, img_np: np.ndarray, task_id: str = "segment"):
h, w = img_np.shape[:2]
print(f"[{task_id}] Running segmentation (backend: {segmentation_backend})...", flush=True)
start_time = time.perf_counter()
if segmentation_backend == "oneformer":
inputs = seg_processor(
images=img,
task_inputs=["semantic"],
return_tensors="pt",
).to(device)
with torch.no_grad():
outputs = seg_model(**inputs)
result = seg_processor.post_process_semantic_segmentation(
outputs,
target_sizes=[(h, w)],
)[0]
seg_map = result.cpu().numpy().astype(np.uint8)
elif segmentation_backend == "mask2former":
inputs = seg_processor(images=img, return_tensors="pt").to(device)
with torch.no_grad():
outputs = seg_model(**inputs)
is_panoptic = "panoptic" in MASK2FORMER_MODEL_NAME
if is_panoptic:
pan_result = seg_processor.post_process_panoptic_segmentation(
outputs,
target_sizes=[(h, w)],
)[0]
seg_map = np.zeros((h, w), dtype=np.uint8)
pan_map = pan_result["segmentation"].cpu().numpy()
for seg_info in pan_result["segments_info"]:
seg_map[pan_map == seg_info["id"]] = min(seg_info["label_id"], 255)
else:
result = seg_processor.post_process_semantic_segmentation(
outputs,
target_sizes=[(h, w)],
)[0]
seg_map = result.cpu().numpy().astype(np.uint8)
else:
inputs = seg_processor(images=img, return_tensors="pt").to(device)
with torch.no_grad():
outputs = seg_model(**inputs)
seg = outputs.logits.argmax(dim=1).squeeze().cpu().numpy()
seg_map = cv2.resize(seg.astype(np.uint8), (w, h), interpolation=cv2.INTER_NEAREST)
duration = time.perf_counter() - start_time
print(f"[{task_id}] Segmentation completed in {duration:.4f}s", flush=True)
return seg_map
def segmenter_metadata_name() -> str:
if segmentation_backend == "oneformer":
return "oneformer-ade20k-swin-large"
return segmentation_backend
def build_segmentation_bundle(contents: bytes, task_id: str = "segment"):
print(f"[{task_id}] Starting bundle build...", flush=True)
start_total = time.perf_counter()
# Load original high-resolution image. Bake in the EXIF orientation so the
# pixels are upright before any processing — phone photos (e.g. iPhone) are
# stored in sensor orientation with an EXIF rotation tag, which we re-encode
# away below, so without this the output bundle would come out rotated.
img = Image.open(io.BytesIO(contents))
img = ImageOps.exif_transpose(img).convert("RGB")
img_np = np.array(img)
h, w = img_np.shape[:2]
# Calculate scale factor and downscaled image
max_dim = 1024
if w > max_dim or h > max_dim:
if w > h:
new_w = max_dim
new_h = int(round(h * max_dim / w))
else:
new_h = max_dim
new_w = int(round(w * max_dim / h))
img_resized = img.resize((new_w, new_h), Image.Resampling.BILINEAR)
img_np_resized = np.array(img_resized)
else:
img_resized = img
img_np_resized = img_np
new_w, new_h = w, h
# Run the full processing pipeline at the downscaled resolution for high speed
min_floor_area_resized = max(1200, int(new_w * new_h * 0.015))
seg_map_resized = run_segmentation(img_resized, img_np_resized, task_id=task_id)
primary_floor_ids = class_ids(PRIMARY_FLOOR_CLASSES)
floor_class_ids = class_ids(FLOOR_SURFACE_CLASSES)
floor_mask_resized = np.isin(seg_map_resized, primary_floor_ids).astype(np.uint8)
floor_mask_resized = wall_subtract(floor_mask_resized, seg_map_resized, dilation=1)
floor_mask_resized = clean_floor_mask(floor_mask_resized)
if int(floor_mask_resized.sum()) < min_floor_area_resized:
floor_mask_resized = np.isin(seg_map_resized, floor_class_ids).astype(np.uint8)
floor_mask_resized = wall_subtract(floor_mask_resized, seg_map_resized, dilation=1)
floor_mask_resized = clean_floor_mask(floor_mask_resized)
depth_resized = estimate_depth(img_resized, new_w, new_h, task_id=task_id)
homography_resized, plane_resized = None, None
if FLOOR_PLANE_FROM_DEPTH and depth_resized is not None:
homography_resized, plane_resized = estimate_floor_plane_from_depth(
floor_mask_resized, depth_resized, task_id=task_id
)
if homography_resized is None:
homography_resized, plane_resized = estimate_floor_plane(
floor_mask_resized, img_np_resized, task_id=task_id
)
quad_resized = np.asarray(plane_resized["quad"], dtype=np.float32).reshape(4, 2) if plane_resized and plane_resized.get("quad") else None
surface_mask_resized = build_floor_surface_mask(floor_mask_resized, seg_map_resized, quad_resized, depth_resized, task_id=task_id)
shade_map_resized = None
if ENABLE_INTRINSIC_SHADING and intrinsic_models is not None:
shade_map_resized = build_intrinsic_shade_map(img_np_resized, surface_mask_resized, task_id=task_id)
if shade_map_resized is None:
shade_map_resized = build_shade_map(img_np_resized, surface_mask_resized, task_id=task_id)
# Now upscale the results back to the original high-resolution space
# 1. Base64-encode the original high-resolution image as a high-quality JPEG
buffer = io.BytesIO()
img.save(buffer, format="JPEG", quality=95)
pixels_b64 = base64.b64encode(buffer.getvalue()).decode()
# 2. Resize masks back to original resolution
seg_map = cv2.resize(seg_map_resized, (w, h), interpolation=cv2.INTER_NEAREST)
floor_mask = cv2.resize(floor_mask_resized, (w, h), interpolation=cv2.INTER_NEAREST)
surface_mask = cv2.resize(surface_mask_resized, (w, h), interpolation=cv2.INTER_NEAREST)
surface_indices = np.flatnonzero(surface_mask.ravel()).astype(np.uint32)
if depth_resized is not None:
depth = cv2.resize(depth_resized, (w, h), interpolation=cv2.INTER_LINEAR)
else:
depth = None
if shade_map_resized is not None:
shade_map = cv2.resize(shade_map_resized, (w, h), interpolation=cv2.INTER_LINEAR)
else:
shade_map = None
# 3. Scale homography and plane coordinates to original resolution
sx, sy = w / new_w, h / new_h
homography = None
plane = None
if homography_resized is not None:
# The homography maps source image pixels -> destination floor-plane
# coords, both estimated in the downscaled space. To express it in the
# original-resolution space we conjugate by the scale transform:
# H_orig = diag(sx, sy, 1) @ H_resized @ diag(1/sx, 1/sy, 1)
# The right factor rescales the input pixels (columns); the left factor
# rescales the output floor coords (rows) so they match the plane rect,
# which is also scaled to original resolution below.
h_matrix = np.array(homography_resized).reshape(3, 3)
h_matrix[:, 0] /= sx
h_matrix[:, 1] /= sy
h_matrix[0, :] *= sx
h_matrix[1, :] *= sy
homography = h_matrix.flatten().tolist()
if plane_resized is not None:
# Scale quad coords
quad_orig = []
if plane_resized.get("quad"):
for i, val in enumerate(plane_resized["quad"]):
if i % 2 == 0:
quad_orig.append(val * sx)
else:
quad_orig.append(val * sy)
# Scale vanishing point
vp_orig = None
if plane_resized.get("vanishingPoint") is not None:
vp_orig = {
"x": plane_resized["vanishingPoint"]["x"] * sx,
"y": plane_resized["vanishingPoint"]["y"] * sy,
}
plane = {
"x": plane_resized["x"] * sx,
"y": plane_resized["y"] * sy,
"width": plane_resized["width"] * sx,
"height": plane_resized["height"] * sy,
"quad": quad_orig,
"vanishingPoint": vp_orig,
}
# Populate final segments bundle list
min_floor_area = max(1200, int(w * h * 0.015))
segments = []
if len(surface_indices) >= min_floor_area:
segments.append({
"id": 0,
"className": "floor",
"mask": base64.b64encode(surface_indices.tobytes()).decode(),
"homography": homography,
"plane": plane,
"shadeMap": base64.b64encode(shade_map.tobytes()).decode() if shade_map is not None else None,
"metadata": {
"segmenter": segmenter_metadata_name(),
"floorPixels": int(floor_mask.sum()),
"surfacePixels": int(surface_mask.sum()),
"depthEnabled": depth is not None,
"shadingEnabled": shade_map is not None,
},
})
if not segments:
flat_seg = seg_map.ravel()
for seg_id, class_id in enumerate(np.unique(flat_seg)):
indices = np.where(flat_seg == class_id)[0].astype(np.uint32)
if len(indices) < 1000:
continue
segments.append({
"id": int(seg_id),
"className": class_name_for_id(int(class_id)),
"mask": base64.b64encode(indices.tobytes()).decode(),
"homography": None,
"plane": None,
"shadeMap": None,
"metadata": {
"segmenter": segmenter_metadata_name(),
"depthEnabled": depth is not None,
"shadingEnabled": False,
},
})
duration = time.perf_counter() - start_total
print(f"[{task_id}] Bundle build completed in {duration:.4f}s", flush=True)
return {"width": w, "height": h, "pixels": pixels_b64, "segments": segments}
def job_path(job_id: str) -> Path:
return JOB_DIR / f"{job_id}.json"
def read_job(job_id: str):
path = job_path(job_id)
if not path.exists():
raise HTTPException(status_code=404, detail="Job not found.")
return json.loads(path.read_text())
def write_job(job: dict):
job_path(job["id"]).write_text(json.dumps(job))
def run_conversion_task(job_id: str, upload_path: Path):
print(f"[{job_id}] Starting background conversion task...", flush=True)
start_time = time.perf_counter()
try:
image_bytes = upload_path.read_bytes()
bundle = build_segmentation_bundle(image_bytes, task_id=job_id)
(JOB_DIR / f"{job_id}.bundle.json").write_text(json.dumps(bundle))
job = read_job(job_id)
job["status"] = "COMPLETED"
write_job(job)
duration = time.perf_counter() - start_time
print(f"[{job_id}] Background conversion task completed in {duration:.4f}s", flush=True)
except Exception as exc:
duration = time.perf_counter() - start_time
print(f"[{job_id}] Background conversion failed after {duration:.4f}s: {exc}", flush=True)
try:
job = read_job(job_id)
job["status"] = "FAILED"
job["error"] = str(exc)
write_job(job)
except Exception:
pass
@app.get("/")
async def root():
return Response(status_code=200)
@app.post("/viz2d/convert")
async def convert_to_viz2d(file: UploadFile = File(...)):
if file.content_type and not file.content_type.startswith("image/"):
raise HTTPException(status_code=400, detail="Upload must be a JPG or PNG image.")
job_id = uuid.uuid4().hex
ext = Path(file.filename or "room.jpg").suffix.lower()
if ext not in {".jpg", ".jpeg", ".png", ".webp"}:
ext = ".jpg"
upload_path = UPLOAD_DIR / f"{job_id}{ext}"
with upload_path.open("wb") as out:
shutil.copyfileobj(file.file, out)
job = {
"id": job_id,
"status": "PROCESSING",
"inputUrl": f"/uploads/{upload_path.name}",
"outputUrl": f"/viz2d/jobs/{job_id}/file",
}
write_job(job)
# Queue the work on the serialized inference pool rather than Starlette's
# background-task thread pool, which would run many jobs against the shared
# model at once. The job is fire-and-forget; status is tracked on disk.
INFERENCE_POOL.submit(run_conversion_task, job_id, upload_path)
return job
@app.get("/viz2d/jobs/{job_id}")
async def viz2d_job_status(job_id: str):
return read_job(job_id)
@app.get("/viz2d/jobs/{job_id}/file")
async def viz2d_job_file(job_id: str):
job = read_job(job_id)
if job.get("status") != "COMPLETED":
raise HTTPException(status_code=409, detail="Job is not completed yet.")
bundle_path = JOB_DIR / f"{job_id}.bundle.json"
if not bundle_path.exists():
raise HTTPException(status_code=404, detail="Job output not found.")
return Response(
content=bundle_path.read_bytes(),
media_type="application/json",
headers={"Content-Disposition": 'attachment; filename="visualizer.vizbundle.json"'},
)
@app.post("/segment")
async def segment(file: UploadFile = File(...)):
contents = await file.read()
task_id = f"segment_{uuid.uuid4().hex[:8]}"
# Run the heavy inference in the serialized pool so it never blocks the event
# loop (status polls, health checks) and never overlaps another inference.
loop = asyncio.get_running_loop()
return await loop.run_in_executor(
INFERENCE_POOL, build_segmentation_bundle, contents, task_id
)
if __name__ == "__main__":
import uvicorn
uvicorn.run(app, host="0.0.0.0", port=7860)