Spaces:
Running
on
Zero
Running
on
Zero
remove padding
Browse files- visualization/et_visualizer.py +20 -53
visualization/et_visualizer.py
CHANGED
@@ -12,37 +12,18 @@ from visualization.logger import SimulationLogger
|
|
12 |
from scipy.spatial.transform import Rotation
|
13 |
|
14 |
|
15 |
-
def load_trajectory_data(traj_file: str, char_file: str
|
16 |
trajectory = read_kitti_poses_file(traj_file)
|
17 |
matrix_trajectory = torch.from_numpy(
|
18 |
np.array(trajectory.poses_se3)).to(torch.float32)
|
19 |
|
20 |
-
raw_trans = torch.clone(matrix_trajectory[:, :3, 3])
|
21 |
-
raw_rot = matrix_trajectory[:, :3, :3]
|
22 |
-
|
23 |
-
rot6d = raw_rot[:, :, :2].permute(0, 2, 1).reshape(-1, 6)
|
24 |
-
trajectory_feature = torch.hstack([rot6d, raw_trans]).permute(1, 0)
|
25 |
-
|
26 |
-
padded_trajectory_feature = F.pad(
|
27 |
-
trajectory_feature,
|
28 |
-
(0, num_cams - trajectory_feature.shape[1])
|
29 |
-
)
|
30 |
-
|
31 |
-
padding_mask = torch.ones((num_cams))
|
32 |
-
padding_mask[trajectory_feature.shape[1]:] = 0
|
33 |
-
|
34 |
char_feature = torch.from_numpy(np.load(char_file)).to(torch.float32)
|
35 |
-
padding_size = num_cams - char_feature.shape[0]
|
36 |
-
padded_char_feature = F.pad(
|
37 |
-
char_feature, (0, 0, 0, padding_size)).permute(1, 0)
|
38 |
|
39 |
return {
|
40 |
"traj_filename": Path(traj_file).name,
|
41 |
"char_filename": Path(char_file).name,
|
42 |
-
"
|
43 |
-
"
|
44 |
-
"padding_mask": padding_mask,
|
45 |
-
"raw_matrix_trajectory": matrix_trajectory
|
46 |
}
|
47 |
|
48 |
|
@@ -58,11 +39,8 @@ class ETLogger(SimulationLogger):
|
|
58 |
[0, 0, 1]
|
59 |
])
|
60 |
|
61 |
-
def log_trajectory(self, trajectory: np.ndarray
|
62 |
-
|
63 |
-
valid_trajectory = trajectory[:valid_frames]
|
64 |
-
|
65 |
-
positions = valid_trajectory[:, :3, 3]
|
66 |
rr.log(
|
67 |
"world/trajectory/points",
|
68 |
rr.Points3D(
|
@@ -83,13 +61,12 @@ class ETLogger(SimulationLogger):
|
|
83 |
timeless=True
|
84 |
)
|
85 |
|
86 |
-
for k in range(
|
87 |
-
|
88 |
rr.set_time_sequence("frame_idx", k)
|
89 |
|
90 |
-
translation =
|
91 |
rotation_q = Rotation.from_matrix(
|
92 |
-
|
93 |
|
94 |
rr.log(
|
95 |
f"world/camera",
|
@@ -108,20 +85,16 @@ class ETLogger(SimulationLogger):
|
|
108 |
),
|
109 |
)
|
110 |
|
111 |
-
def log_character(self, char_feature: np.ndarray
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
(valid_char.reshape(-1, 3).shape[0], 4), [0.8, 0.2, 0.2, 1.0])
|
122 |
-
),
|
123 |
-
timeless=True
|
124 |
-
)
|
125 |
|
126 |
|
127 |
@spaces.GPU
|
@@ -134,14 +107,8 @@ def visualize_et_data(traj_file: str, char_file: str) -> Optional[str]:
|
|
134 |
rrd_path = os.path.join(temp_dir, "et_visualization.rrd")
|
135 |
|
136 |
logger = ETLogger()
|
137 |
-
logger.log_trajectory(
|
138 |
-
|
139 |
-
data["padding_mask"].numpy()
|
140 |
-
)
|
141 |
-
logger.log_character(
|
142 |
-
data["char_feat"].numpy(),
|
143 |
-
data["padding_mask"].numpy()
|
144 |
-
)
|
145 |
|
146 |
rr.save(rrd_path)
|
147 |
return rrd_path
|
|
|
12 |
from scipy.spatial.transform import Rotation
|
13 |
|
14 |
|
15 |
+
def load_trajectory_data(traj_file: str, char_file: str) -> Dict:
|
16 |
trajectory = read_kitti_poses_file(traj_file)
|
17 |
matrix_trajectory = torch.from_numpy(
|
18 |
np.array(trajectory.poses_se3)).to(torch.float32)
|
19 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
20 |
char_feature = torch.from_numpy(np.load(char_file)).to(torch.float32)
|
|
|
|
|
|
|
21 |
|
22 |
return {
|
23 |
"traj_filename": Path(traj_file).name,
|
24 |
"char_filename": Path(char_file).name,
|
25 |
+
"char_feat": char_feature,
|
26 |
+
"matrix_trajectory": matrix_trajectory
|
|
|
|
|
27 |
}
|
28 |
|
29 |
|
|
|
39 |
[0, 0, 1]
|
40 |
])
|
41 |
|
42 |
+
def log_trajectory(self, trajectory: np.ndarray):
|
43 |
+
positions = trajectory[:, :3, 3]
|
|
|
|
|
|
|
44 |
rr.log(
|
45 |
"world/trajectory/points",
|
46 |
rr.Points3D(
|
|
|
61 |
timeless=True
|
62 |
)
|
63 |
|
64 |
+
for k in range(len(trajectory)):
|
|
|
65 |
rr.set_time_sequence("frame_idx", k)
|
66 |
|
67 |
+
translation = trajectory[k, :3, 3]
|
68 |
rotation_q = Rotation.from_matrix(
|
69 |
+
trajectory[k, :3, :3]).as_quat()
|
70 |
|
71 |
rr.log(
|
72 |
f"world/camera",
|
|
|
85 |
),
|
86 |
)
|
87 |
|
88 |
+
def log_character(self, char_feature: np.ndarray):
|
89 |
+
rr.log(
|
90 |
+
"world/character",
|
91 |
+
rr.Points3D(
|
92 |
+
char_feature.reshape(-1, 3),
|
93 |
+
colors=np.full(
|
94 |
+
(char_feature.reshape(-1, 3).shape[0], 4), [0.8, 0.2, 0.2, 1.0])
|
95 |
+
),
|
96 |
+
timeless=True
|
97 |
+
)
|
|
|
|
|
|
|
|
|
98 |
|
99 |
|
100 |
@spaces.GPU
|
|
|
107 |
rrd_path = os.path.join(temp_dir, "et_visualization.rrd")
|
108 |
|
109 |
logger = ETLogger()
|
110 |
+
logger.log_trajectory(data["matrix_trajectory"].numpy())
|
111 |
+
logger.log_character(data["char_feat"].numpy())
|
|
|
|
|
|
|
|
|
|
|
|
|
112 |
|
113 |
rr.save(rrd_path)
|
114 |
return rrd_path
|