File size: 8,648 Bytes
570b87b | 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 | """Phase 3 — vision tower: project boxes/rays onto the HOA-7 sphere.
No neural net. Detections become soft spherical Gaussians in Ambix SN3D
space (same 64-D geometry as audio). Suitable as conditioning / tool output
for language models and later diffusion control.
"""
from __future__ import annotations
import math
from typing import Any, Iterable, Mapping, Optional, Sequence
import numpy as np
from .basis import MAX_ORDER, N_CHANNELS, sh_sn3d, unit_vector
from .encode import mix
from .report import SpatialReport, report_from_hoa
def _as_box(b: Mapping[str, Any]) -> dict:
"""Normalize a box/ray dict."""
if "az" in b or "azimuth" in b or "azimuth_deg" in b:
az = float(b.get("az", b.get("azimuth", b.get("azimuth_deg", 0.0))))
el = float(b.get("el", b.get("elevation", b.get("elevation_deg", 0.0))))
elif "x" in b and "y" in b:
# normalized image coords: x in [0,1] left→right, y in [0,1] top→bottom
# map to az in [-180,180] around center, el in [-90,90]
x = float(b["x"])
y = float(b["y"])
# simple equirectangular pinhole-ish: center (0.5,0.5) = front
# az: left of image = +az (left), right = -az if we want Ambix left=+Y
# Convention: image x=0 left → az=+hfov/2, x=1 right → az=-hfov/2
hfov = float(b.get("hfov_deg", 90.0))
vfov = float(b.get("vfov_deg", 60.0))
az = (0.5 - x) * hfov
el = (0.5 - y) * vfov
else:
raise ValueError(f"box needs az/el or x/y: {b}")
return {
"az": az,
"el": el,
"w_deg": float(b.get("w_deg", b.get("width_deg", b.get("sigma_deg", 8.0)))),
"h_deg": float(b.get("h_deg", b.get("height_deg", b.get("sigma_deg", 8.0)))),
"weight": float(b.get("weight", b.get("score", b.get("confidence", 1.0)))),
"label": str(b.get("label", b.get("class", b.get("name", "")))),
"kind": str(b.get("kind", "box")), # box | ray | point
}
def angular_gaussian_weights(
center_az: float,
center_el: float,
sample_az: np.ndarray,
sample_el: np.ndarray,
sigma_az_deg: float,
sigma_el_deg: float,
) -> np.ndarray:
"""Soft lobe on the sphere (product of wrapped azimuth + elevation Gaussians)."""
# great-circle-ish separation using unit vectors is better
c = unit_vector(center_az, center_el, degrees=True)
# sample_az, sample_el may be mesh
u = unit_vector(sample_az, sample_el, degrees=True)
# cos gamma = c · u
cosg = np.clip(np.sum(u * c, axis=-1), -1.0, 1.0)
gamma = np.rad2deg(np.arccos(cosg))
# isotropic sigma from geometric mean of width axes
sig = max(0.5, math.sqrt(max(sigma_az_deg, 0.5) * max(sigma_el_deg, 0.5)))
return np.exp(-0.5 * (gamma / sig) ** 2)
def encode_boxes_to_hoa(
boxes: Sequence[Mapping[str, Any]],
*,
max_order: int = MAX_ORDER,
n_azi: int = 72,
n_el: int = 36,
) -> np.ndarray:
"""Project one or more vision boxes/rays into a static HOA-7 field (C,).
Each box is a soft spherical Gaussian; the field is the weighted sum of
SN3D plane-wave encodings of the discrete sphere samples of that lobe.
"""
if not boxes:
return np.zeros((max_order + 1) ** 2, dtype=np.float64)
azi = np.linspace(-180.0, 180.0, n_azi, endpoint=False)
el = np.linspace(-90.0, 90.0, n_el)
AA, EE = np.meshgrid(azi, el, indexing="ij")
# solid-angle-ish weights
w_el = np.cos(np.deg2rad(el))
w_el = np.maximum(w_el, 0.0)
dA = (2.0 * math.pi / n_azi) * (math.pi / max(n_el - 1, 1))
sa = dA * w_el[None, :] # (1,E) broadcast to (A,E)
field = np.zeros((max_order + 1) ** 2, dtype=np.float64)
for raw in boxes:
b = _as_box(raw)
if b["kind"] == "ray" or b["kind"] == "point":
# delta: single direction plane wave
Y = sh_sn3d(b["az"], b["el"], degrees=True, max_order=max_order)
field = field + b["weight"] * Y
continue
lobe = angular_gaussian_weights(
b["az"], b["el"], AA, EE, b["w_deg"], b["h_deg"]
)
# discrete HOA analysis of the lobe function f(Ω)
# a = sum f(Ω) Y(Ω) sa(Ω) * (2n+1)/4π — use same scale as rotate dense
Y = sh_sn3d(AA, EE, degrees=True, max_order=max_order) # (A,E,C)
f = lobe * b["weight"]
# a_c = sum_{ae} f_ae * sa_ae * Y_ae,c
weighted = (f * sa)[..., None] * Y
a = np.sum(weighted, axis=(0, 1))
# SN3D analysis scale per order
nch = a.shape[0]
scale = np.empty(nch, dtype=np.float64)
for n in range(max_order + 1):
s = float(2 * n + 1)
for m in range(-n, n + 1):
scale[n * (n + 1) + m] = s
a = a * scale / (4.0 * math.pi)
field = field + a
# pad to 64
out = np.zeros(N_CHANNELS, dtype=np.float64)
out[: field.shape[0]] = field
return out
def report_from_boxes(
boxes: Sequence[Mapping[str, Any]],
*,
max_order: int = MAX_ORDER,
sample_rate: int = 48000,
n_samples: int = 1,
meta: Optional[dict] = None,
) -> SpatialReport:
"""Vision-only spatial report from detection boxes/rays."""
hoa = encode_boxes_to_hoa(boxes, max_order=max_order)
# report_from_hoa expects (C,T) — use a constant "frame"
stream = np.tile(hoa.reshape(-1, 1), (1, max(1, n_samples)))
notes = [f"vision tower: {len(boxes)} box/ray event(s) on HOA sphere"]
hints = []
for i, raw in enumerate(boxes):
try:
b = _as_box(raw)
hints.append(
{
"label": b["label"] or f"box{i}",
"azimuth_deg": b["az"],
"elevation_deg": b["el"],
"weight": b["weight"],
"kind": b["kind"],
}
)
except Exception:
continue
m = dict(meta or {})
m["encode"] = "vision_boxes"
m["n_boxes"] = len(boxes)
rep = report_from_hoa(
stream,
sample_rate,
max_order=max_order,
include_frames=False,
include_bands=False,
include_peak_map=True,
notes=notes,
meta=m,
)
rep.kind = "spatial_vision"
rep.sources_hint = hints
return rep
def fuse_reports(audio: Mapping[str, Any], vision: Mapping[str, Any]) -> dict:
"""Merge audio + vision spatial reports into one agent-facing payload.
Strategy:
- Keep both DOAs
- angular_separation_deg between them
- agreement flag if within 15°
- combined one_liner
"""
from .analysis import angular_error_deg
a_az = float(audio.get("doa_az_deg", 0.0))
a_el = float(audio.get("doa_el_deg", 0.0))
v_az = float(vision.get("doa_az_deg", vision.get("peak_az_deg", 0.0)))
v_el = float(vision.get("doa_el_deg", vision.get("peak_el_deg", 0.0)))
sep = angular_error_deg(a_az, a_el, v_az, v_el)
agree = sep <= 15.0
# energy-weighted blend of unit vectors if both have energy
from .basis import az_el_from_unit, unit_vector
ea = float(audio.get("energy", 0.0)) + 1e-12
ev = float(vision.get("energy", 0.0)) + 1e-12
ua = unit_vector(a_az, a_el, degrees=True)
uv = unit_vector(v_az, v_el, degrees=True)
um = (ea * ua + ev * uv) / (ea + ev)
nm = float(np.linalg.norm(um))
if nm > 1e-15:
baz, bel = az_el_from_unit(um / nm, degrees=True)
baz, bel = float(baz), float(bel)
else:
baz, bel = a_az, a_el
one = (
f"spatial-fuse: audio=({a_az:.1f},{a_el:.1f}) vision=({v_az:.1f},{v_el:.1f}) "
f"sep={sep:.1f}° agree={agree} blend=({baz:.1f},{bel:.1f})"
)
return {
"schema": "spatial-hoa.fuse.v1",
"kind": "spatial_av_fuse",
"audio_doa_az_deg": a_az,
"audio_doa_el_deg": a_el,
"vision_doa_az_deg": v_az,
"vision_doa_el_deg": v_el,
"angular_separation_deg": sep,
"agreement": agree,
"blend_az_deg": baz,
"blend_el_deg": bel,
"audio_energy": float(audio.get("energy", 0.0)),
"vision_energy": float(vision.get("energy", 0.0)),
"audio_one_liner": audio.get("one_liner")
or f"audio DOA ({a_az:.1f},{a_el:.1f})",
"vision_one_liner": vision.get("one_liner")
or f"vision DOA ({v_az:.1f},{v_el:.1f})",
"vision_sources": vision.get("sources_hint") or [],
"audio_bands": audio.get("bands") or [],
"one_liner": one,
"notes": [
"Fused audio intensity DOA with vision spherical presence peak.",
"agreement true if separation ≤ 15°.",
],
}
|