File size: 6,924 Bytes
fadb92b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import native_rasterizer
import torch
import torch.nn as nn
from torch.autograd import Function

MODE_BOUNDARY = "boundary"
MODE_MASK = "mask"
MODE_HARD_MASK = "hard_mask"

MODE_MAPPING = {MODE_BOUNDARY: 0, MODE_MASK: 1, MODE_HARD_MASK: 2}


class SoftPolygonFunction(Function):
    @staticmethod
    def forward(ctx, vertices, width, height, inv_smoothness=1.0, mode=MODE_BOUNDARY):
        ctx.width = width
        ctx.height = height
        ctx.inv_smoothness = inv_smoothness
        ctx.mode = MODE_MAPPING[mode]

        vertices = vertices.clone()
        ctx.device = vertices.device
        ctx.batch_size, ctx.number_vertices = vertices.shape[:2]

        rasterized = torch.FloatTensor(ctx.batch_size, ctx.height, ctx.width).fill_(0.0).to(device=ctx.device)

        contribution_map = torch.IntTensor(ctx.batch_size, ctx.height, ctx.width).fill_(0).to(device=ctx.device)
        rasterized, contribution_map = native_rasterizer.forward_rasterize(
            vertices, rasterized, contribution_map, width, height, inv_smoothness, ctx.mode
        )
        ctx.save_for_backward(vertices, rasterized, contribution_map)

        return rasterized  # , contribution_map

    @staticmethod
    def backward(ctx, grad_output):
        vertices, rasterized, contribution_map = ctx.saved_tensors

        grad_output = grad_output.contiguous()

        # grad_vertices = torch.FloatTensor(
        #    ctx.batch_size, ctx.height, ctx.width, ctx.number_vertices, 2).fill_(0.0).to(device=ctx.device)
        grad_vertices = torch.FloatTensor(ctx.batch_size, ctx.number_vertices, 2).fill_(0.0).to(device=ctx.device)
        grad_vertices = native_rasterizer.backward_rasterize(
            vertices,
            rasterized,
            contribution_map,
            grad_output,
            grad_vertices,
            ctx.width,
            ctx.height,
            ctx.inv_smoothness,
            ctx.mode,
        )

        return grad_vertices, None, None, None, None


class SoftPolygon(nn.Module):
    MODES = [MODE_BOUNDARY, MODE_MASK, MODE_HARD_MASK]

    def __init__(self, inv_smoothness=1.0, mode=MODE_BOUNDARY):
        super(SoftPolygon, self).__init__()

        self.inv_smoothness = inv_smoothness

        if mode not in SoftPolygon.MODES:
            raise ValueError("invalid mode: {0}".format(mode))

        self.mode = mode

    def forward(self, vertices, width, height, p, color=False):
        return SoftPolygonFunction.apply(vertices, width, height, self.inv_smoothness, self.mode)


def pnp(vertices, width, height):
    device = vertices.device
    batch_size = vertices.size(0)
    polygon_dimension = vertices.size(1)

    y_index = torch.arange(0, height).to(device)
    x_index = torch.arange(0, width).to(device)

    grid_y, grid_x = torch.meshgrid(y_index, x_index)
    xp = grid_x.unsqueeze(0).repeat(batch_size, 1, 1).float()
    yp = grid_y.unsqueeze(0).repeat(batch_size, 1, 1).float()

    result = torch.zeros((batch_size, height, width)).bool().to(device)

    j = polygon_dimension - 1
    for vn in range(polygon_dimension):
        from_x = vertices[:, vn, 0].unsqueeze(-1).unsqueeze(-1).repeat(1, height, width)
        from_y = vertices[:, vn, 1].unsqueeze(-1).unsqueeze(-1).repeat(1, height, width)

        to_x = vertices[:, j, 0].unsqueeze(-1).unsqueeze(-1).repeat(1, height, width)
        to_y = vertices[:, j, 1].unsqueeze(-1).unsqueeze(-1).repeat(1, height, width)

        has_condition = torch.logical_and(
            (from_y > yp) != (to_y > yp), xp < ((to_x - from_x) * (yp - from_y) / (to_y - from_y) + from_x)
        )

        if has_condition.any():
            result[has_condition] = ~result[has_condition]

        j = vn

    signed_result = -torch.ones((batch_size, height, width), device=device)
    signed_result[result] = 1.0

    return signed_result


# used for verification purposes only.
class SoftPolygonPyTorch(nn.Module):
    def __init__(self, inv_smoothness=1.0):
        super(SoftPolygonPyTorch, self).__init__()

        self.inv_smoothness = inv_smoothness

    # vertices is N x P x 2
    def forward(self, vertices, width, height, p, color=False):
        device = vertices.device
        batch_size = vertices.size(0)
        polygon_dimension = vertices.size(1)

        inside_outside = pnp(vertices, width, height)

        # discrete points we will sample from.
        y_index = torch.arange(0, height).to(device)
        x_index = torch.arange(0, width).to(device)

        grid_y, grid_x = torch.meshgrid(y_index, x_index)
        grid_x = grid_x.unsqueeze(0).repeat(batch_size, 1, 1).float()
        grid_y = grid_y.unsqueeze(0).repeat(batch_size, 1, 1).float()

        # do this "per dimension"
        distance_segments = []
        over_segments = []
        for from_index in range(polygon_dimension):
            segment_result = torch.zeros((batch_size, height, width)).to(device)
            from_vertex = vertices[:, from_index].unsqueeze(-1).unsqueeze(-1)

            if from_index == (polygon_dimension - 1):
                to_vertex = vertices[:, 0].unsqueeze(-1).unsqueeze(-1)
            else:
                to_vertex = vertices[:, from_index + 1].unsqueeze(-1).unsqueeze(-1)

            x2_sub_x1 = to_vertex[:, 0] - from_vertex[:, 0]
            y2_sub_y1 = to_vertex[:, 1] - from_vertex[:, 1]
            square_segment_length = x2_sub_x1 * x2_sub_x1 + y2_sub_y1 * y2_sub_y1 + 0.00001

            # figure out if this is a major/minor segment (todo?)
            x_sub_x1 = grid_x - from_vertex[:, 0]
            y_sub_y1 = grid_y - from_vertex[:, 1]
            x_sub_x2 = grid_x - to_vertex[:, 0]
            y_sub_y2 = grid_y - to_vertex[:, 1]

            # dot between the given point and first vertex and first vertex and second vertex.
            dot = ((x_sub_x1 * x2_sub_x1) + (y_sub_y1 * y2_sub_y1)) / square_segment_length

            # needlessly computed sometimes.
            x_proj = grid_x - (from_vertex[:, 0] + dot * x2_sub_x1)
            y_proj = grid_y - (from_vertex[:, 1] + dot * y2_sub_y1)

            from_closest = dot < 0
            to_closest = dot > 1
            interior_closest = (dot >= 0) & (dot <= 1)

            segment_result[from_closest] = x_sub_x1[from_closest] ** 2 + y_sub_y1[from_closest] ** 2
            segment_result[to_closest] = x_sub_x2[to_closest] ** 2 + y_sub_y2[to_closest] ** 2
            segment_result[interior_closest] = x_proj[interior_closest] ** 2 + y_proj[interior_closest] ** 2

            distance_map = -segment_result
            distance_segments.append(distance_map)

            signed_map = torch.sigmoid(-distance_map * inside_outside / self.inv_smoothness)
            over_segments.append(signed_map)

        F_max, F_arg = torch.max(torch.stack(distance_segments, dim=-1), dim=-1)
        F_theta = torch.gather(torch.stack(over_segments, dim=-1), dim=-1, index=F_arg.unsqueeze(-1))[..., 0]

        return F_theta