-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathevalution.py
205 lines (167 loc) · 7.98 KB
/
evalution.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
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
import os
import cv2
import torch
import numpy as np
from tqdm import tqdm
from torch.utils.data import DataLoader
from pytorch3d.ops import ball_query, knn_points
from pytorch3d.ops.utils import masked_gather
from diff_gaussian_rasterization.scene.cameras import PerspectiveCamera
from utils.render import render
from utils.metrics import eval_metrics
from utils.image import render_semantic
from models.loss import MESMaskedLoss
def mse2psnr(mse):
"""
:param mse: scalar
:return: scalar np.float32
"""
mse = np.maximum(mse, 1e-10) # avoid -inf or nan when mse is very small.
psnr = -10.0 * np.log10(mse)
return psnr.astype(np.float32)
def eval_metric(gaussians, exposure_model, dataset, bg_color, pipe, eval_root=None):
loss_all = []
loss_fuction = MESMaskedLoss()
image_segs = []
gt_segs = []
cnt = 0
num_class = dataset.num_class
device = "cuda:0"
with torch.no_grad():
dataloader = DataLoader(dataset, batch_size=1, num_workers=4, shuffle=True, drop_last=True)
for sample in tqdm(dataloader):
for key, value in sample.items():
if key != "image_name":
sample[key] = value[0].to(device) # 注意batch_size
else:
sample[key] = value[0]
gt_image = sample["image"]
mask = sample["mask"]
gt_seg = sample["label"]
image_idx = sample["idx"].item()
cam_idx = sample["cam_idx"].item()
R = sample["R"]
T = sample["T"]
image_name = sample["image_name"]
viewpoint_cam = PerspectiveCamera(R, T, sample["K"], sample["W"], sample["H"], 1, 10, device)
render_pkg = render(viewpoint_cam, gaussians, pipe, bg_color)
src_render_image, render_depth, valid_mask = render_pkg["render"], render_pkg["depth"], render_pkg["mask"]
render_image = exposure_model(cam_idx, src_render_image)
render_image = render_image.permute(1, 2, 0)
mask = mask * valid_mask.float()
label_feature = render(viewpoint_cam, gaussians, pipe, bg_color, render_type="label")
render_seg = label_feature["render"]
render_seg = render_seg.permute(1, 2, 0)
mse_loss = loss_fuction(render_image, gt_image, mask[..., None])
mse_loss_np = mse_loss.cpu().detach().numpy()
loss_all.append(mse_loss_np)
render_image = render_image.detach().cpu().numpy()
gt_image = gt_image.detach().cpu().numpy()
render_image = (render_image * 255).astype(np.uint8)[:, :, ::-1] # RGB2BGR
gt_image = (gt_image * 255).astype(np.uint8)[:, :, ::-1] # RGB2BGR
if eval_root:
cv2.imwrite(os.path.join(eval_root, f"{image_name}-render.png"), render_image)
cv2.imwrite(os.path.join(eval_root, f"{image_name}.png"), gt_image)
# save seg numpy array
images_seg_np = render_seg.detach().cpu().numpy()[None] # (1, H, W, C)
images_seg_np = np.argmax(images_seg_np, axis=-1) # (1, H, W)
vis_seg = render_semantic(images_seg_np[0], dataset.filted_color_map)[:, :, ::-1]
blend_image = cv2.addWeighted(gt_image, 0.5, vis_seg, 0.5, 0)
if eval_root:
cv2.imwrite(os.path.join(eval_root, f"{image_name}-vis_seg.png"), vis_seg)
cv2.imwrite(os.path.join(eval_root, f"{image_name}-blend.png"), blend_image)
images_seg_np[images_seg_np == num_class - 1] = 255
images_seg_np[images_seg_np == 0] = 255
images_seg_np -= 1
images_seg_np[images_seg_np == 254] = 255
image_segs.append(images_seg_np)
mask = mask.detach().cpu().numpy().astype(np.uint8)
gt_seg_np = gt_seg.detach().cpu().numpy()
gt_seg_np *= mask
gt_seg_np = gt_seg_np[None]
vis_gt_seg = render_semantic(gt_seg_np[0], dataset.filted_color_map)[:, :, ::-1]
if eval_root:
cv2.imwrite(os.path.join(eval_root, f"{image_name}-vis_gt_seg.png"), vis_gt_seg)
gt_seg_np[gt_seg_np == num_class - 1] = 255
gt_seg_np[gt_seg_np == 0] = 255
gt_seg_np -= 1
gt_seg_np[gt_seg_np == 254] = 255
gt_segs.append(gt_seg_np)
cnt += 1
loss_all = np.array(loss_all)
loss_mean = np.mean(loss_all)
psnr_mean = mse2psnr(loss_mean)
if len(image_segs) > 1:
image_segs = np.concatenate(image_segs, axis=0)
gt_segs = np.concatenate(gt_segs, axis=0)
else:
image_segs = np.array(image_segs)[None]
gt_segs = np.array(gt_segs)[None]
results = eval_metrics(image_segs, gt_segs,
num_classes=num_class - 2,
ignore_index=255,
metrics=['mIoU'],
nan_to_num=None,
label_map=dict(),
reduce_zero_label=False)
return {"MSE": loss_mean, "PSNR": psnr_mean, "mIoU": results}
def eval_bev_metric(gt_root, pre_root, num_class):
mask1 = cv2.imread(os.path.join(gt_root, "bev_mask.png"), cv2.IMREAD_GRAYSCALE)
mask1 = mask1 > 0
mask2 = cv2.imread(os.path.join(pre_root, "bev_mask.png"), cv2.IMREAD_GRAYSCALE)
mask2 = mask2 > 0
mask = mask1 * mask2 # (H,W)
# ======> seg metric
gt_label = cv2.imread(os.path.join(gt_root, "bev_label.png"), cv2.IMREAD_GRAYSCALE).astype(np.uint8) # (H,W)
pre_label = cv2.imread(os.path.join(pre_root, "bev_label.png"), cv2.IMREAD_GRAYSCALE).astype(np.uint8) # (H,W)
gt_label *= mask
gt_label = gt_label[None]
gt_label[gt_label == num_class - 1] = 255
gt_label[gt_label == 0] = 255
gt_label -= 1
gt_label[gt_label == 254] = 255
pre_label = pre_label[None]
pre_label[pre_label == num_class - 1] = 255
pre_label[pre_label == 0] = 255
pre_label -= 1
pre_label[pre_label == 254] = 255
pre_label = pre_label[None]
gt_label = gt_label[None]
results = eval_metrics(pre_label, gt_label,
num_classes=num_class - 2,
ignore_index=255,
metrics=['mIoU'],
nan_to_num=None,
label_map=dict(),
reduce_zero_label=False)
# =====> rgb metric
loss_fuction = MESMaskedLoss()
gt_image = cv2.imread(os.path.join(gt_root, "bev_image.png"), cv2.IMREAD_COLOR) # (H,W,3)
render_image = cv2.imread(os.path.join(pre_root, "bev_image.png"), cv2.IMREAD_COLOR) # (H,W,3)
gt_image = gt_image / 255.0
render_image = render_image / 255.0
mse_loss = loss_fuction(torch.from_numpy(render_image), torch.from_numpy(gt_image), torch.from_numpy(mask[..., None]))
mse_loss_np = mse_loss.numpy()
loss_mean = np.mean(mse_loss_np)
psnr_mean = mse2psnr(loss_mean)
return {"MSE": loss_mean, "PSNR": psnr_mean, "mIoU": results}
def eval_z_metric(lidar_xyz, guassian_xyz):
near_idx = ball_query(guassian_xyz[None, :, :2], lidar_xyz[None, :, :2], K=1, return_nn=False, radius=0.1).idx # (1, N, 1)
valid_mask = near_idx.squeeze(0) != -1 # (N, 1)
gt_z = masked_gather(lidar_xyz[None, :, 2:3], near_idx).reshape(-1, 1) # (N, 1)
gt_z = gt_z[valid_mask] # (M, 1)
pre_z = guassian_xyz[:, 2:3][valid_mask] # (M, 1)
loss = torch.sqrt(torch.mean((gt_z - pre_z) ** 2))
return loss.item()
def eval_chamfer_metric(lidar_xyz, guassian_xyz):
dists1 = knn_points(guassian_xyz[None], lidar_xyz[None], K=1, return_nn=False).dists
dists2 = knn_points(lidar_xyz[None], guassian_xyz[None], K=1, return_nn=False).dists
# 排序
dists1 = dists1.flatten()
dists2 = dists2.flatten()
dist1 = torch.sort(dists1)[0]
dist1 = dist1[:int(0.97 * len(dist1))]
dist2 = torch.sort(dists2)[0]
dist2 = dist2[:int(0.97 * len(dist2))]
eval_chamfer_metric = torch.sqrt(torch.mean(dist1 ** 2)) + torch.sqrt(torch.mean(dist2 ** 2))
return eval_chamfer_metric.item()