File size: 7,860 Bytes
efb793f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import numpy as np
import trimesh
from matplotlib import colormaps
from scipy.spatial.transform import Rotation


def predictions_to_glb(
    predictions: dict,
    conf_thres: float = 50.0,
    filter_by_frames: str = "All",
    remove_far_points: bool = True,
    static_only: bool = False,
    show_cam: bool = True,
    prediction_mode: str = "Depthmap and Camera Branch",
) -> trimesh.Scene:
    """Convert Wat3R predictions to a GLB scene."""
    if not isinstance(predictions, dict):
        raise TypeError("predictions must be a dictionary")

    points, confidence = _select_points(predictions, prediction_mode)
    images = _images_to_nhwc(predictions["images"])
    depth = _depth_to_shw(predictions.get("depth"))
    static_mask = _mask_to_shw(predictions.get("static_mask"))
    extrinsics = predictions["extrinsic"]

    frame_index = _parse_frame_index(filter_by_frames)
    if frame_index is not None:
        points = points[frame_index][None]
        confidence = confidence[frame_index][None]
        images = images[frame_index][None]
        if depth is not None:
            depth = depth[frame_index][None]
        if static_mask is not None:
            static_mask = static_mask[frame_index][None]
        extrinsics = extrinsics[frame_index][None]

    vertices = points.reshape(-1, 3)
    colors = (images.reshape(-1, 3).clip(0, 1) * 255).astype(np.uint8)
    confidence = confidence.reshape(-1)

    if conf_thres == 0:
        conf_threshold = 0.0
    else:
        conf_threshold = np.percentile(confidence, conf_thres)
    valid_mask = (confidence >= conf_threshold) & (confidence > 1e-5) & np.isfinite(vertices).all(axis=1)

    if remove_far_points and depth is not None:
        valid_mask &= _two_mean_near_mask(depth).reshape(-1)
    if static_only:
        if static_mask is None:
            valid_mask &= False
        else:
            valid_mask &= static_mask.reshape(-1)

    vertices = vertices[valid_mask]
    colors = colors[valid_mask]
    scene_scale = _scene_scale(vertices)

    scene = trimesh.Scene()
    scene.add_geometry(trimesh.PointCloud(vertices=vertices, colors=colors))

    extrinsics_4x4 = _to_4x4(extrinsics)
    if show_cam:
        cmap = colormaps.get_cmap("gist_rainbow")
        for index, world_to_camera in enumerate(extrinsics_4x4):
            camera_to_world = np.linalg.inv(world_to_camera)
            rgba = cmap(index / max(1, len(extrinsics_4x4)))
            color = tuple(int(255 * value) for value in rgba[:3])
            _add_camera(scene, camera_to_world, color, scene_scale)

    if len(extrinsics_4x4) > 0:
        _align_scene_to_first_camera(scene, extrinsics_4x4[0])
    return scene


def _select_points(predictions: dict, prediction_mode: str):
    if "Pointmap" in prediction_mode and "world_points" in predictions:
        return (
            predictions["world_points"],
            predictions.get("world_points_conf", np.ones_like(predictions["world_points"][..., 0])),
        )
    return (
        predictions["world_points_from_depth"],
        predictions.get("depth_conf", np.ones_like(predictions["world_points_from_depth"][..., 0])),
    )


def _images_to_nhwc(images: np.ndarray) -> np.ndarray:
    if images.ndim != 4:
        raise ValueError(f"Expected images with shape [S, C, H, W] or [S, H, W, C], got {images.shape}")
    if images.shape[1] == 3:
        images = np.transpose(images, (0, 2, 3, 1))
    return images


def _depth_to_shw(depth: np.ndarray | None) -> np.ndarray | None:
    if depth is None:
        return None
    depth = np.asarray(depth)
    if depth.ndim == 5 and depth.shape[0] == 1:
        depth = depth[0]
    if depth.ndim == 4 and depth.shape[-1] == 1:
        depth = depth[..., 0]
    if depth.ndim != 3:
        return None
    return depth


def _mask_to_shw(mask: np.ndarray | None) -> np.ndarray | None:
    if mask is None:
        return None
    mask = np.asarray(mask)
    if mask.ndim == 5 and mask.shape[0] == 1:
        mask = mask[0]
    if mask.ndim == 4 and mask.shape[-1] == 1:
        mask = mask[..., 0]
    if mask.ndim != 3:
        return None
    return mask.astype(bool)


def _two_mean_near_mask(
    depth: np.ndarray,
    num_iters: int = 20,
    clip_percentiles: tuple[float, float] = (1.0, 99.0),
) -> np.ndarray:
    """Split each depth map into near/far clusters and keep the near cluster."""
    depth = np.asarray(depth, dtype=np.float32)
    masks = []

    for frame_depth in depth:
        finite = np.isfinite(frame_depth)
        if not finite.any():
            masks.append(np.zeros(frame_depth.shape, dtype=bool))
            continue

        values = frame_depth[finite]
        low, high = np.percentile(values, clip_percentiles)
        values_clipped = np.clip(values, low, high)

        center_1 = float(values_clipped.min())
        center_2 = float(values_clipped.max())
        if abs(center_1 - center_2) < 1e-8:
            masks.append(finite)
            continue

        for _ in range(num_iters):
            assign_1 = np.abs(values_clipped - center_1) <= np.abs(values_clipped - center_2)
            if assign_1.any():
                center_1 = float(values_clipped[assign_1].mean())
            if (~assign_1).any():
                center_2 = float(values_clipped[~assign_1].mean())

        near_center = min(center_1, center_2)
        far_center = max(center_1, center_2)
        threshold = 0.5 * (near_center + far_center) * 0.9
        mask = finite & (frame_depth < threshold)
        masks.append(mask if mask.any() else finite)

    return np.stack(masks, axis=0)


def _parse_frame_index(frame_filter: str):
    if frame_filter in (None, "All", "all"):
        return None
    try:
        return int(str(frame_filter).split(":")[0])
    except (TypeError, ValueError, IndexError):
        return None


def _to_4x4(extrinsics: np.ndarray) -> np.ndarray:
    extrinsics = np.asarray(extrinsics)
    if extrinsics.ndim != 3 or extrinsics.shape[1:] != (3, 4):
        raise ValueError(f"Expected extrinsics with shape [S, 3, 4], got {extrinsics.shape}")
    matrices = np.tile(np.eye(4, dtype=np.float32), (extrinsics.shape[0], 1, 1))
    matrices[:, :3, :4] = extrinsics
    return matrices


def _scene_scale(vertices: np.ndarray) -> float:
    if vertices.size == 0:
        return 1.0
    lower = np.percentile(vertices, 5, axis=0)
    upper = np.percentile(vertices, 95, axis=0)
    scale = float(np.linalg.norm(upper - lower))
    return scale if scale > 1e-6 else 1.0


def _add_camera(scene: trimesh.Scene, camera_to_world: np.ndarray, color: tuple[int, int, int], scene_scale: float) -> None:
    cam_width = scene_scale * 0.05
    cam_height = scene_scale * 0.1

    rotate_45 = np.eye(4)
    rotate_45[:3, :3] = Rotation.from_euler("z", 45, degrees=True).as_matrix()
    rotate_45[2, 3] = -cam_height

    transform = camera_to_world @ _opencv_to_opengl() @ rotate_45
    cone = trimesh.creation.cone(cam_width, cam_height, sections=4)
    vertices = _transform_points(transform, cone.vertices)
    faces = cone.faces
    mesh = trimesh.Trimesh(vertices=vertices, faces=faces)
    mesh.visual.face_colors[:, :3] = color
    scene.add_geometry(mesh)


def _align_scene_to_first_camera(scene: trimesh.Scene, first_world_to_camera: np.ndarray) -> None:
    align_rotation = np.eye(4)
    align_rotation[:3, :3] = Rotation.from_euler("y", 180, degrees=True).as_matrix()
    scene.apply_transform(np.linalg.inv(first_world_to_camera) @ _opencv_to_opengl() @ align_rotation)


def _opencv_to_opengl() -> np.ndarray:
    matrix = np.eye(4)
    matrix[1, 1] = -1
    matrix[2, 2] = -1
    return matrix


def _transform_points(transform: np.ndarray, points: np.ndarray) -> np.ndarray:
    points_h = np.concatenate([points, np.ones((points.shape[0], 1))], axis=1)
    return (points_h @ transform.T)[:, :3]