File size: 6,405 Bytes
2df809d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os.path as osp
import os
import sys
import itertools

sys.path.append(osp.join(osp.dirname(__file__), "..", ".."))
import cv2
import numpy as np
import h5py
import math
from dust3r.datasets.base.base_multiview_dataset import BaseMultiViewDataset
from dust3r.utils.image import imread_cv2


class ARKitScenesHighRes_Multi(BaseMultiViewDataset):
    def __init__(self, *args, split, ROOT, **kwargs):
        self.ROOT = ROOT
        self.video = True
        self.max_interval = 8
        self.is_metric = True
        super().__init__(*args, **kwargs)
        if split == "train":
            self.split = "Training"
        elif split == "test":
            self.split = "Validation"
        else:
            raise ValueError("")

        self.loaded_data = self._load_data(self.split)

    def _load_data(self, split):
        all_scenes = sorted(
            [
                d
                for d in os.listdir(osp.join(self.ROOT, split))
                if osp.isdir(osp.join(self.ROOT, split, d))
            ]
        )
        offset = 0
        scenes = []
        sceneids = []
        images = []
        start_img_ids = []
        scene_img_list = []
        timestamps = []
        intrinsics = []
        trajectories = []
        scene_id = 0
        for scene in all_scenes:
            scene_dir = osp.join(self.ROOT, self.split, scene)
            with np.load(osp.join(scene_dir, "scene_metadata.npz")) as data:
                imgs_with_indices = sorted(
                    enumerate(data["images"]), key=lambda x: x[1]
                )
                imgs = [x[1] for x in imgs_with_indices]
                cut_off = (
                    self.num_views
                    if not self.allow_repeat
                    else max(self.num_views // 3, 3)
                )
                if len(imgs) < cut_off:
                    print(f"Skipping {scene}")
                    continue
                indices = [x[0] for x in imgs_with_indices]
                tsps = np.array(
                    [float(img_name.split("_")[1][:-4]) for img_name in imgs]
                )
                assert [img[:8] == scene for img in imgs], f"{scene}, {imgs}"
                num_imgs = data["images"].shape[0]
                img_ids = list(np.arange(num_imgs) + offset)
                start_img_ids_ = img_ids[: num_imgs - cut_off + 1]

                scenes.append(scene)
                scene_img_list.append(img_ids)
                sceneids.extend([scene_id] * num_imgs)
                images.extend(imgs)
                start_img_ids.extend(start_img_ids_)
                timestamps.extend(tsps)

                K = np.expand_dims(np.eye(3), 0).repeat(num_imgs, 0)
                intrins = data["intrinsics"][indices]
                K[:, 0, 0] = [fx for _, _, fx, _, _, _ in intrins]
                K[:, 1, 1] = [fy for _, _, _, fy, _, _ in intrins]
                K[:, 0, 2] = [cx for _, _, _, _, cx, _ in intrins]
                K[:, 1, 2] = [cy for _, _, _, _, _, cy in intrins]
                intrinsics.extend(list(K))
                trajectories.extend(list(data["trajectories"][indices]))

                # offset groups
                offset += num_imgs
                scene_id += 1

        self.scenes = scenes
        self.sceneids = sceneids
        self.images = images
        self.scene_img_list = scene_img_list
        self.intrinsics = intrinsics
        self.trajectories = trajectories
        self.start_img_ids = start_img_ids
        assert len(self.images) == len(self.intrinsics) == len(self.trajectories)

    def __len__(self):
        return len(self.start_img_ids)

    def get_image_num(self):
        return len(self.images)

    def _get_views(self, idx, resolution, rng, num_views):
        start_id = self.start_img_ids[idx]
        all_image_ids = self.scene_img_list[self.sceneids[start_id]]
        pos, ordered_video = self.get_seq_from_start_id(
            num_views,
            start_id,
            all_image_ids,
            rng,
            max_interval=self.max_interval,
            block_shuffle=16,
        )
        image_idxs = np.array(all_image_ids)[pos]

        views = []

        for v, view_idx in enumerate(image_idxs):
            scene_id = self.sceneids[view_idx]
            scene_dir = osp.join(self.ROOT, self.split, self.scenes[scene_id])

            intrinsics = self.intrinsics[view_idx]
            camera_pose = self.trajectories[view_idx]
            basename = self.images[view_idx]
            assert (
                basename[:8] == self.scenes[scene_id]
            ), f"{basename}, {self.scenes[scene_id]}"
            # print(scene_dir, basename)
            # Load RGB image
            rgb_image = imread_cv2(
                osp.join(scene_dir, "vga_wide", basename.replace(".png", ".jpg"))
            )
            # Load depthmap
            depthmap = imread_cv2(
                osp.join(scene_dir, "highres_depth", basename), cv2.IMREAD_UNCHANGED
            )
            depthmap = depthmap.astype(np.float32) / 1000.0
            depthmap[~np.isfinite(depthmap)] = 0  # invalid

            rgb_image, depthmap, intrinsics = self._crop_resize_if_necessary(
                rgb_image, depthmap, intrinsics, resolution, rng=rng, info=view_idx
            )

            # generate img mask and raymap mask
            img_mask, ray_mask = self.get_img_and_ray_masks(
                self.is_metric, v, rng, p=[0.7, 0.25, 0.05]
            )

            views.append(
                dict(
                    img=rgb_image,
                    depthmap=depthmap.astype(np.float32),
                    camera_pose=camera_pose.astype(np.float32),
                    camera_intrinsics=intrinsics.astype(np.float32),
                    dataset="arkitscenes_highres",
                    label=self.scenes[scene_id] + "_" + basename,
                    instance=f"{str(idx)}_{str(view_idx)}",
                    is_metric=self.is_metric,
                    is_video=ordered_video,
                    quantile=np.array(0.99, dtype=np.float32),
                    img_mask=img_mask,
                    ray_mask=ray_mask,
                    camera_only=False,
                    depth_only=False,
                    single_view=False,
                    reset=False,
                )
            )
        assert len(views) == num_views
        return views