-
Notifications
You must be signed in to change notification settings - Fork 195
/
predictor.py
175 lines (152 loc) · 6.61 KB
/
predictor.py
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
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import torch
import torch.nn.functional as F
from tqdm import tqdm
from cotracker.models.core.cotracker.cotracker import get_points_on_a_grid
from cotracker.models.core.model_utils import smart_cat
from cotracker.models.build_cotracker import (
build_cotracker,
)
class CoTrackerPredictor(torch.nn.Module):
def __init__(
self, checkpoint="cotracker/checkpoints/cotracker_stride_4_wind_8.pth"
):
super().__init__()
self.interp_shape = (384, 512)
self.support_grid_size = 6
model = build_cotracker(checkpoint)
self.model = model
self.model.eval()
@torch.no_grad()
def forward(
self,
video, # (1, T, 3, H, W)
# input prompt types:
# - None. Dense tracks are computed in this case. You can adjust *query_frame* to compute tracks starting from a specific frame.
# *backward_tracking=True* will compute tracks in both directions.
# - queries. Queried points of shape (1, N, 3) in format (t, x, y) for frame index and pixel coordinates.
# - grid_size. Grid of N*N points from the first frame. if segm_mask is provided, then computed only for the mask.
# You can adjust *query_frame* and *backward_tracking* for the regular grid in the same way as for dense tracks.
queries: torch.Tensor = None,
segm_mask: torch.Tensor = None, # Segmentation mask of shape (B, 1, H, W)
grid_size: int = 0,
grid_query_frame: int = 0, # only for dense and regular grid tracks
backward_tracking: bool = False,
):
if queries is None and grid_size == 0:
tracks, visibilities = self._compute_dense_tracks(
video,
grid_query_frame=grid_query_frame,
backward_tracking=backward_tracking,
)
else:
tracks, visibilities = self._compute_sparse_tracks(
video,
queries,
segm_mask,
grid_size,
add_support_grid=(grid_size == 0 or segm_mask is not None),
grid_query_frame=grid_query_frame,
backward_tracking=backward_tracking,
)
return tracks, visibilities
def _compute_dense_tracks(
self, video, grid_query_frame, grid_size=30, backward_tracking=False
):
*_, H, W = video.shape
grid_step = W // grid_size
grid_width = W // grid_step
grid_height = H // grid_step
tracks = visibilities = None
grid_pts = torch.zeros((1, grid_width * grid_height, 3)).to(video.device)
grid_pts[0, :, 0] = grid_query_frame
for offset in tqdm(range(grid_step * grid_step)):
ox = offset % grid_step
oy = offset // grid_step
grid_pts[0, :, 1] = (
torch.arange(grid_width).repeat(grid_height) * grid_step + ox
)
grid_pts[0, :, 2] = (
torch.arange(grid_height).repeat_interleave(grid_width) * grid_step + oy
)
tracks_step, visibilities_step = self._compute_sparse_tracks(
video=video,
queries=grid_pts,
backward_tracking=backward_tracking,
)
tracks = smart_cat(tracks, tracks_step, dim=2)
visibilities = smart_cat(visibilities, visibilities_step, dim=2)
return tracks, visibilities
def _compute_sparse_tracks(
self,
video,
queries,
segm_mask=None,
grid_size=0,
add_support_grid=False,
grid_query_frame=0,
backward_tracking=False,
):
B, T, C, H, W = video.shape
assert B == 1
video = video.reshape(B * T, C, H, W)
video = F.interpolate(video, tuple(self.interp_shape), mode="bilinear")
video = video.reshape(B, T, 3, self.interp_shape[0], self.interp_shape[1])
if queries is not None:
queries = queries.clone()
B, N, D = queries.shape
assert D == 3
queries[:, :, 1] *= self.interp_shape[1] / W
queries[:, :, 2] *= self.interp_shape[0] / H
elif grid_size > 0:
grid_pts = get_points_on_a_grid(grid_size, self.interp_shape, device=video.device)
if segm_mask is not None:
segm_mask = F.interpolate(
segm_mask, tuple(self.interp_shape), mode="nearest"
)
point_mask = segm_mask[0, 0][
(grid_pts[0, :, 1]).round().long().cpu(),
(grid_pts[0, :, 0]).round().long().cpu(),
].bool()
grid_pts = grid_pts[:, point_mask]
queries = torch.cat(
[torch.ones_like(grid_pts[:, :, :1]) * grid_query_frame, grid_pts],
dim=2,
)
if add_support_grid:
grid_pts = get_points_on_a_grid(self.support_grid_size, self.interp_shape, device=video.device)
grid_pts = torch.cat(
[torch.zeros_like(grid_pts[:, :, :1]), grid_pts], dim=2
)
queries = torch.cat([queries, grid_pts], dim=1)
tracks, __, visibilities, __ = self.model(rgbs=video, queries=queries, iters=6)
if backward_tracking:
tracks, visibilities = self._compute_backward_tracks(
video, queries, tracks, visibilities
)
if add_support_grid:
queries[:, -self.support_grid_size ** 2 :, 0] = T - 1
if add_support_grid:
tracks = tracks[:, :, : -self.support_grid_size ** 2]
visibilities = visibilities[:, :, : -self.support_grid_size ** 2]
thr = 0.9
visibilities = visibilities > thr
tracks[:, :, :, 0] *= W / float(self.interp_shape[1])
tracks[:, :, :, 1] *= H / float(self.interp_shape[0])
return tracks, visibilities
def _compute_backward_tracks(self, video, queries, tracks, visibilities):
inv_video = video.flip(1).clone()
inv_queries = queries.clone()
inv_queries[:, :, 0] = inv_video.shape[1] - inv_queries[:, :, 0] - 1
inv_tracks, __, inv_visibilities, __ = self.model(
rgbs=inv_video, queries=inv_queries, iters=6
)
inv_tracks = inv_tracks.flip(1)
inv_visibilities = inv_visibilities.flip(1)
mask = tracks == 0
tracks[mask] = inv_tracks[mask]
visibilities[mask[:, :, :, 0]] = inv_visibilities[mask[:, :, :, 0]]
return tracks, visibilities