File size: 5,413 Bytes
55e58d1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import numpy as np
from pytorch3d.ops import ball_query
import torch
import open3d as o3d
from utils.geometry import denoise
from torch.nn.utils.rnn import pad_sequence

COVERAGE_THRESHOLD = 0.3
DISTANCE_THRESHOLD = 0.03
FEW_POINTS_THRESHOLD = 25
DEPTH_TRUNC = 20
BBOX_EXPAND = 0.1


def backproject(depth, intrinisc_cam_parameters, extrinsics):
    """
    convert color and depth to view pointcloud
    """
    depth = o3d.geometry.Image(depth)
    pcld = o3d.geometry.PointCloud.create_from_depth_image(depth, intrinisc_cam_parameters, depth_scale=1, depth_trunc=DEPTH_TRUNC)
    pcld.transform(extrinsics)
    return pcld


def get_neighbor(valid_points, scene_points, lengths_1, lengths_2):
    _, neighbor_in_scene_pcld, _ = ball_query(valid_points, scene_points, lengths_1, lengths_2, K=20, radius=DISTANCE_THRESHOLD, return_nn=False)
    return neighbor_in_scene_pcld


def get_depth_mask(depth):
    depth_tensor = torch.from_numpy(depth).cuda()
    depth_mask = torch.logical_and(depth_tensor > 0, depth_tensor < DEPTH_TRUNC).reshape(-1)
    return depth_mask


def crop_scene_points(mask_points, scene_points):
    x_min, x_max = torch.min(mask_points[:, 0]), torch.max(mask_points[:, 0])
    y_min, y_max = torch.min(mask_points[:, 1]), torch.max(mask_points[:, 1])
    z_min, z_max = torch.min(mask_points[:, 2]), torch.max(mask_points[:, 2])

    selected_point_mask = (scene_points[:, 0] > x_min) & (scene_points[:, 0] < x_max) & (scene_points[:, 1] > y_min) & (scene_points[:, 1] < y_max) & (scene_points[:, 2] > z_min) & (scene_points[:, 2] < z_max)
    selected_point_ids = torch.where(selected_point_mask)[0]
    cropped_scene_points = scene_points[selected_point_ids]
    return cropped_scene_points, selected_point_ids


def turn_mask_to_point(dataset, scene_points, mask_image, frame_id):
    intrinisc_cam_parameters = dataset.get_intrinsics(frame_id)
    extrinsics = dataset.get_extrinsic(frame_id)
    if np.sum(np.isinf(extrinsics)) > 0:
        return {}, [], set()

    depth = dataset.get_depth(frame_id)
    depth_mask = get_depth_mask(depth)
    
    mask_image = torch.from_numpy(mask_image).cuda().reshape(-1)
    ids = torch.unique(mask_image).cpu().numpy()
    ids.sort()
    

    colored_pcld = backproject(depth, intrinisc_cam_parameters, extrinsics)
    view_points = np.asarray(colored_pcld.points)

    mask_points_list = []
    mask_points_num_list = []
    scene_points_list = []
    scene_points_num_list = []
    selected_point_ids_list = []
    initial_valid_mask_ids = []
    for mask_id in ids:
        if mask_id == 0:
            continue
        segmentation = mask_image == mask_id
        valid_mask = segmentation[depth_mask].cpu().numpy()

        mask_pcld = o3d.geometry.PointCloud()
        try:
            mask_points = view_points[valid_mask]
        except IndexError:
            print(f"Error in mask_id: {mask_id}")
            continue
        if len(mask_points) < FEW_POINTS_THRESHOLD:
            continue
        mask_pcld.points = o3d.utility.Vector3dVector(mask_points)

        mask_pcld = mask_pcld.voxel_down_sample(voxel_size=DISTANCE_THRESHOLD)
        mask_pcld, _ = denoise(mask_pcld)
        mask_points = np.asarray(mask_pcld.points)
        if len(mask_points) < FEW_POINTS_THRESHOLD:
            continue
        
        mask_points = torch.tensor(mask_points).float().cuda()
        cropped_scene_points, selected_point_ids = crop_scene_points(mask_points, scene_points)
        initial_valid_mask_ids.append(mask_id)
        mask_points_list.append(mask_points)
        scene_points_list.append(cropped_scene_points)
        mask_points_num_list.append(len(mask_points))
        scene_points_num_list.append(len(cropped_scene_points))
        selected_point_ids_list.append(selected_point_ids)

    if len(initial_valid_mask_ids) == 0:
        return {}, [], []
    mask_points_tensor = pad_sequence(mask_points_list, batch_first=True, padding_value=0)
    scene_points_tensor = pad_sequence(scene_points_list, batch_first=True, padding_value=0)

    lengths_1 = torch.tensor(mask_points_num_list).cuda()
    lengths_2 = torch.tensor(scene_points_num_list).cuda()
    neighbor_in_scene_pcld = get_neighbor(mask_points_tensor, scene_points_tensor, lengths_1, lengths_2)

    valid_mask_ids = []
    mask_info = {}
    frame_point_ids = set()

    for i, mask_id in enumerate(initial_valid_mask_ids):
        mask_neighbor = neighbor_in_scene_pcld[i] # P, 20
        mask_point_num = mask_points_num_list[i] # Pi
        mask_neighbor = mask_neighbor[:mask_point_num] # Pi, 20

        valid_neighbor = mask_neighbor != -1 # Pi, 20
        neighbor = torch.unique(mask_neighbor[valid_neighbor])
        neighbor_in_complete_scene_points = selected_point_ids_list[i][neighbor].cpu().numpy()
        coverage = torch.any(valid_neighbor, dim=1).sum().item() / mask_point_num

        if coverage < COVERAGE_THRESHOLD:
            continue
        valid_mask_ids.append(mask_id)
        mask_info[mask_id] = set(neighbor_in_complete_scene_points)
        frame_point_ids.update(mask_info[mask_id])

    return mask_info, valid_mask_ids, list(frame_point_ids)


def frame_backprojection(dataset, scene_points, frame_id):
    mask_image = dataset.get_segmentation(frame_id, align_with_depth=True)
    mask_info, _, frame_point_ids = turn_mask_to_point(dataset, scene_points, mask_image, frame_id)
    return mask_info, frame_point_ids