Image Segmentation
PyTorch
English
painting-vision-robotics-kit
semantic-segmentation
robotics
edge-ai
construction-ai
autonomous-painting
wall-painting-robot
paint-coverage-estimation
building-facade
drywall
skirting-detection
window-detection
lidar
depth-validation
segformer
deeplabv3
Eval Results (legacy)
Upload test_depth_fusion.py with huggingface_hub
Browse files- test_depth_fusion.py +130 -0
test_depth_fusion.py
ADDED
|
@@ -0,0 +1,130 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Tests for depth_fusion.py on depth rendered from known geometry. Run: python3 test_depth_fusion.py
|
| 3 |
+
|
| 4 |
+
Scenes are the camera-projected rooms of test_wall_measure.py (feet in the world),
|
| 5 |
+
with a depth map ray-cast against the true wall, ceiling, floor and side-wall
|
| 6 |
+
planes, then corrupted the way a phone LiDAR / ToF sensor is: 5 mm noise, 3%
|
| 7 |
+
dropouts, and glass that returns the scene behind it.
|
| 8 |
+
"""
|
| 9 |
+
import numpy as np
|
| 10 |
+
|
| 11 |
+
import test_wall_measure as tw
|
| 12 |
+
from depth_fusion import measure_with_depth
|
| 13 |
+
from wall_measure import PAINTED, SKIRTING, WINDOW
|
| 14 |
+
|
| 15 |
+
FT = 0.3048
|
| 16 |
+
K = (tw.F, tw.F, tw.W / 2, tw.H / 2)
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
def render_depth(cam, planes, noise_m=0.005, dropout=0.03, seed=0):
|
| 20 |
+
"""Depth (metres, OpenCV z) of the nearest of ``planes`` = [(normal, offset_ft, bounds)] per pixel."""
|
| 21 |
+
r, c = cam
|
| 22 |
+
yy, xx = np.mgrid[0:tw.H, 0:tw.W] + 0.0
|
| 23 |
+
d_cam = np.stack([(xx - tw.W / 2) / tw.F, -(yy - tw.H / 2) / tw.F, np.ones_like(xx)], -1) # test camera: y up
|
| 24 |
+
d_world = d_cam @ r # R^T d for row vectors
|
| 25 |
+
best = np.full((tw.H, tw.W), np.inf)
|
| 26 |
+
for normal, offset, bounds in planes:
|
| 27 |
+
normal = np.asarray(normal, float)
|
| 28 |
+
denom = d_world @ normal
|
| 29 |
+
with np.errstate(divide="ignore", invalid="ignore"):
|
| 30 |
+
t = -(c @ normal + offset) / denom
|
| 31 |
+
hit = c + d_world * t[..., None]
|
| 32 |
+
ok = (t > 0) & np.isfinite(t)
|
| 33 |
+
for axis, (lo, hi) in bounds.items():
|
| 34 |
+
ok &= (hit[..., axis] >= lo - 1e-6) & (hit[..., axis] <= hi + 1e-6)
|
| 35 |
+
best = np.where(ok & (t < best), t, best)
|
| 36 |
+
depth = best * FT # z component of d_cam is 1
|
| 37 |
+
rng = np.random.default_rng(seed)
|
| 38 |
+
depth = depth + rng.normal(0, noise_m, depth.shape)
|
| 39 |
+
depth[rng.random(depth.shape) < dropout] = 0.0
|
| 40 |
+
return depth
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
ROOM = [((0, 0, 1), 0.0, {0: (-30, 30), 1: (0, 8)}), # back wall z = 0
|
| 44 |
+
((0, 1, 0), -8.0, {0: (-30, 30), 2: (-30, 0)}), # ceiling y = 8
|
| 45 |
+
((0, 1, 0), 0.0, {0: (-30, 30), 2: (-30, 0)})] # floor y = 0
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
def test_single_wall_to_the_centimetre():
|
| 49 |
+
cam = tw.camera(yaw_deg=-25, pitch_deg=-5, x_offset=1.0, distance=12.0, eye_height=5.0)
|
| 50 |
+
labels = tw.scene(width=12.0, cam=cam)
|
| 51 |
+
# The wall in the labels is 12 ft wide; the room's back plane is wider, so cap it in depth too.
|
| 52 |
+
planes = [((0, 0, 1), 0.0, {0: (0, 12), 1: (0, 8)})] + ROOM[1:]
|
| 53 |
+
depth = render_depth(cam, planes)
|
| 54 |
+
depth[labels == WINDOW] += 2.5 # glass: the sensor sees through it
|
| 55 |
+
m = measure_with_depth(labels, depth, K)
|
| 56 |
+
assert m["status"] == "measured" and len(m["faces"]) == 1, m.get("reason")
|
| 57 |
+
face = m["faces"][0]
|
| 58 |
+
assert face["stop_line_trust"] == {"floor": "depth_planes", "top": "depth_planes", "bottom": "depth_planes"}
|
| 59 |
+
assert abs(face["width_mm"] - 12 * 304.8) < 20, face["width_mm"] # 12 ft within 2 cm
|
| 60 |
+
assert abs(face["height_mm"] - 8 * 304.8) < 10, face["height_mm"] # 8 ft within 1 cm
|
| 61 |
+
assert abs(face["skirting_height_in"] - 4.0) < 0.5, face["skirting_height_in"]
|
| 62 |
+
assert abs(face["area_ft2"]["openings"] - 12.0) < 0.3, face["area_ft2"]
|
| 63 |
+
assert face["plane"]["rms_mm"] < 8 and face["scale_source"] == "depth"
|
| 64 |
+
# Camera-frame stop line: the floor line lies 5 ft below the camera (eye height), pitched.
|
| 65 |
+
floor = np.array(face["stop_lines_camera_mm"]["floor"])
|
| 66 |
+
assert floor.shape == (2, 3)
|
| 67 |
+
|
| 68 |
+
|
| 69 |
+
def test_inside_corner_from_two_planes():
|
| 70 |
+
cam = tw.camera(yaw_deg=-38, x_offset=3.0, distance=9.0, eye_height=4.5)
|
| 71 |
+
back = tw.project(tw.rect(0, 10, 0, 8), cam)
|
| 72 |
+
side = tw.project([(10, 8, 0), (10, 8, -6), (10, 0, -6), (10, 0, 0)], cam)
|
| 73 |
+
labels = np.zeros((tw.H, tw.W), np.uint8)
|
| 74 |
+
tw.fill(labels, back, PAINTED)
|
| 75 |
+
tw.fill(labels, side, PAINTED)
|
| 76 |
+
planes = [((0, 0, 1), 0.0, {0: (-30, 10), 1: (0, 8)}),
|
| 77 |
+
((1, 0, 0), -10.0, {1: (0, 8), 2: (-6, 0)}),
|
| 78 |
+
((0, 1, 0), -8.0, {0: (-30, 10), 2: (-30, 0)}),
|
| 79 |
+
((0, 1, 0), 0.0, {0: (-30, 10), 2: (-30, 0)})]
|
| 80 |
+
depth = render_depth(cam, planes, seed=1)
|
| 81 |
+
m = measure_with_depth(labels, depth, K)
|
| 82 |
+
assert len(m["faces"]) == 2, [f["plane"] for f in m["faces"]]
|
| 83 |
+
back_face, side_face = m["faces"]
|
| 84 |
+
assert back_face["ends"]["right"] == "inside_corner" and side_face["ends"]["left"] == "inside_corner"
|
| 85 |
+
# The corner is the true room corner: both faces stop at the same 3D line.
|
| 86 |
+
a = np.array(back_face["stop_lines_camera_mm"]["right"])
|
| 87 |
+
b = np.array(side_face["stop_lines_camera_mm"]["left"])
|
| 88 |
+
assert np.abs(a - b).max() < 25, (a, b)
|
| 89 |
+
assert abs(side_face["width_mm"] - 6 * 304.8) < 40, side_face["width_mm"] # side wall 6 ft visible
|
| 90 |
+
assert abs(back_face["height_mm"] - 8 * 304.8) < 15
|
| 91 |
+
|
| 92 |
+
|
| 93 |
+
def test_no_ceiling_in_view_falls_back_to_wall_extent():
|
| 94 |
+
cam = tw.camera(yaw_deg=-10, x_offset=6.0, distance=12.0, eye_height=4.0)
|
| 95 |
+
labels = tw.scene(width=12.0, cam=cam, window=None)
|
| 96 |
+
planes = [((0, 0, 1), 0.0, {0: (0, 12), 1: (0, 8)}), ROOM[2]] # floor only
|
| 97 |
+
depth = render_depth(cam, planes, seed=2)
|
| 98 |
+
face = measure_with_depth(labels, depth, K)["faces"][0]
|
| 99 |
+
assert face["stop_line_trust"]["top"] == "depth_extent" and face["stop_line_trust"]["floor"] == "depth_planes"
|
| 100 |
+
assert abs(face["height_mm"] - 8 * 304.8) < 40, face["height_mm"]
|
| 101 |
+
|
| 102 |
+
|
| 103 |
+
def test_too_little_depth_fails_cleanly():
|
| 104 |
+
labels = tw.scene()
|
| 105 |
+
m = measure_with_depth(labels, np.zeros((tw.H, tw.W)), K)
|
| 106 |
+
assert m["status"] == "depth_failed" and m["faces"] == []
|
| 107 |
+
|
| 108 |
+
|
| 109 |
+
def test_low_resolution_depth_is_aligned():
|
| 110 |
+
# Phone LiDAR is 256x192; the RGB frame is larger. Same field of view.
|
| 111 |
+
cam = tw.camera(yaw_deg=-15, x_offset=4.0, distance=12.0, eye_height=4.5)
|
| 112 |
+
labels = tw.scene(width=12.0, cam=cam)
|
| 113 |
+
planes = [((0, 0, 1), 0.0, {0: (0, 12), 1: (0, 8)})] + ROOM[1:]
|
| 114 |
+
depth = render_depth(cam, planes, seed=3)
|
| 115 |
+
small = depth[2::4, 2::4] # 300 x 225, pixel centres preserved
|
| 116 |
+
face = measure_with_depth(labels, small, K)["faces"][0]
|
| 117 |
+
assert abs(face["width_mm"] - 12 * 304.8) < 40, face["width_mm"]
|
| 118 |
+
assert abs(face["height_mm"] - 8 * 304.8) < 20, face["height_mm"]
|
| 119 |
+
|
| 120 |
+
|
| 121 |
+
def main():
|
| 122 |
+
tests = [value for name, value in sorted(globals().items()) if name.startswith("test_")]
|
| 123 |
+
for test in tests:
|
| 124 |
+
test()
|
| 125 |
+
print(f"ok {test.__name__}")
|
| 126 |
+
print(f"{len(tests)} tests passed")
|
| 127 |
+
|
| 128 |
+
|
| 129 |
+
if __name__ == "__main__":
|
| 130 |
+
main()
|