-
Notifications
You must be signed in to change notification settings - Fork 9
/
evaluate_object_reconstruction.py
217 lines (189 loc) · 7.89 KB
/
evaluate_object_reconstruction.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
206
207
208
209
210
211
212
213
214
215
import os,sys
#.path.append("/")
import numpy as np
import trimesh
import argparse
import glob
import pickle as p
import torch
import json
import tempfile
import shutil
import subprocess
from net_utils.bins import *
from external.pyTorchChamferDistance.chamfer_distance import ChamferDistance
import scipy
dist_chamfer=ChamferDistance()
category_label_mapping = {0:"table",
1:"sofa",
2:"cabinet",
3:"night_stand",
4:"chair",
5:"bookshelf",
6:"bed",
7:"desk",
8:"dresser"
}
def parse_args():
'''PARAMETERS'''
parser = argparse.ArgumentParser('totalindoorrecon evaluation')
parser.add_argument('--result_dir', type=str,
help='folder contains the results of object mesh')
parser.add_argument('--gt_dir',type=str,default="/data3/haolin/data/3D-FUTURE-watertight/",help="folder containing the watertight ground truth mesh")
return parser.parse_args()
def delete_disconnected_component(mesh):
split_mesh = mesh.split(only_watertight=False)
max_vertice = 0
max_ind = -1
for idx, mesh in enumerate(split_mesh):
# print(mesh.vertices.shape[0])
if mesh.vertices.shape[0] > max_vertice:
max_vertice = mesh.vertices.shape[0]
max_ind = idx
# print(max_ind)
# print(max_vertice)
return split_mesh[max_ind]
OCCNET_FSCORE_EPS = 1e-09
def percent_below(dists, thresh):
return np.mean((dists**2 <= thresh).astype(np.float32)) * 100.0
def f_score(a_to_b, b_to_a, thresh):
precision = percent_below(a_to_b, thresh)
recall = percent_below(b_to_a, thresh)
return (2 * precision * recall) / (precision + recall + OCCNET_FSCORE_EPS)
def pointcloud_neighbor_distances_indices(source_points, target_points):
target_kdtree = scipy.spatial.cKDTree(target_points)
distances, indices = target_kdtree.query(source_points, n_jobs=-1)
return distances, indices
def fscore(points1,points2,tau=0.002):
"""Computes the F-Score at tau between two meshes."""
dist12, _ = pointcloud_neighbor_distances_indices(points1, points2)
dist21, _ = pointcloud_neighbor_distances_indices(points2, points1)
f_score_tau = f_score(dist12, dist21, tau)
return f_score_tau
args=parse_args()
result_list=glob.glob(args.result_dir+"/*.ply")
prepare_data_dir="./data/3dfront/prepare_data/test"
gt_dir=args.gt_dir
split_path="./data/3dfront/split-filter/test/all.json"
with open(split_path,'r') as f:
split=json.load(f)
select_split_list=[]
for idx,(taskid,object_id) in enumerate(split):
if idx>2000:
break
select_split_list.append((taskid[0:-4],object_id))
def get_rot_from_yaw(yaw):
cy=np.cos(yaw)
sy=np.sin(yaw)
rot=np.array([[cy,0,sy],
[0,1,0],
[-sy,0,cy]])
return rot
chamfer_distance_list=[]
cd_loss_dict={}
fscore_list=[]
fst_dict={}
#select_class_list=["bed"]
log_txt=os.path.join(args.result_dir,"evaluate_log.txt")
for (taskid,object_id) in select_split_list:
#taskid,object_id,depth_error=item
result_file=os.path.join(args.result_dir,"%s_%s.ply"%(taskid,object_id))
#print(result_file)
#print(os.path.isfile(result_file))
if os.path.isfile(result_file)==False:
continue
#print(result_file)
file_name=result_file.split("/")[-1].split(".")[0]
taskid=file_name.split("_")[0]
object_id=file_name.split("_")[1]
#pred_mesh.vertices=np.dot(pred_mesh.vertices,yaw_rot.T)
prepare_data_path=os.path.join(prepare_data_dir,taskid+".pkl")
with open(prepare_data_path,'rb') as f:
prepare_data=p.load(f)
size_cls, size_reg = prepare_data['boxes']['size_cls'][int(object_id)], prepare_data['boxes']['size_reg'][
int(object_id)]
size = avg_size[size_cls] * (1 + size_reg)
classname = category_label_mapping[size_cls]
#if classname not in select_class_list:
# continue
if classname not in cd_loss_dict:
cd_loss_dict[classname]=[]
fst_dict[classname]=[]
jid=prepare_data['boxes']['jid'][int(object_id)]
tran_matrix=prepare_data['boxes']['tran_matrix'][int(object_id)]
cam_center=prepare_data['boxes']['cam_center'][int(object_id)]
K=prepare_data['camera']['K']
wrd2cam_matrix = prepare_data['camera']['wrd2cam_matrix']
rot_matrix = np.dot(wrd2cam_matrix[0:3, 0:3], tran_matrix[0:3, 0:3])
inv_rot=np.linalg.inv(rot_matrix)
#print(prepare_data['boxes'].keys())
try:
pred_mesh = trimesh.load(result_file)
gt_mesh_path=os.path.join(gt_dir,jid,"normalized_watertight.obj")
gt_mesh=trimesh.load(gt_mesh_path)
except:
print(gt_mesh_path,"does not exist")
continue
'''align two mesh firstly'''
#pxmin,pxmax=np.min(pred_mesh.vertices[:,0]),np.max(pred_mesh.vertices[:,0])
#pymin, pymax = np.min(pred_mesh.vertices[:, 1]), np.max(pred_mesh.vertices[:, 1])
#pzmin, pzmax = np.min(pred_mesh.vertices[:, 2]), np.max(pred_mesh.vertices[:, 2])
gxmin, gxmax = np.min(gt_mesh.vertices[:, 0]), np.max(gt_mesh.vertices[:, 0])
gymin, gymax = np.min(gt_mesh.vertices[:, 1]), np.max(gt_mesh.vertices[:, 1])
gzmin, gzmax = np.min(gt_mesh.vertices[:, 2]), np.max(gt_mesh.vertices[:, 2])
#pred_mesh.vertices=pred_mesh.vertices-np.array([(pxmin+pxmax)/2,(pymin+pymax)/2,(pzmin+pzmax)/2])
#pred_mesh.vertices=pred_mesh.vertices/np.array([pxmax-pxmin,pymax-pymin,pzmax-pymin])*2
pred_mesh.vertices=pred_mesh.vertices/2*size/np.max(size)*2
gt_mesh.vertices=gt_mesh.vertices/2*size/np.max(size)*2
temp_folder = tempfile.mktemp(dir='/dev/shm')
os.makedirs(temp_folder)
shutil.copy('./external/ldif/gaps/bin/x86_64/mshalign', temp_folder)
output_file = os.path.join(temp_folder, 'output.ply')
pred_mesh.export(output_file)
align_file = os.path.join(temp_folder, 'align.ply')
gt_file = os.path.join(temp_folder, 'gt.ply')
gt_mesh.export(gt_file)
cmd = f"{os.path.join(temp_folder, 'mshalign')} {output_file} {gt_file} {align_file}"
subprocess.check_output(cmd, shell=True)
align_mesh = trimesh.load(align_file)
#pred_mesh.export("/data1/haolin/alignmesh.ply")
#gt_mesh.export("/data1/haolin/gt_mesh.ply")
cmd="rm -r %s"%(temp_folder)
os.system(cmd)
pred_sample_points=align_mesh.sample(10000)
gt_sample_points=gt_mesh.sample(10000)
fst=fscore(pred_sample_points,gt_sample_points)
fst_dict[classname].append(fst)
fscore_list.append(fst)
pred_sample_gpu=torch.from_numpy(pred_sample_points).float().cuda().unsqueeze(0)
gt_sample_gpu=torch.from_numpy(gt_sample_points).float().cuda().unsqueeze(0)
#print(pred_sample_gpu.shape,gt_sample_gpu.shape)
#loss,_=chamfer_distance(x=pred_sample_gpu,y=gt_sample_gpu)
dist1,dist2=dist_chamfer(gt_sample_gpu,pred_sample_gpu)[:2]
cd_loss=torch.mean(dist1)+torch.mean(dist2)
cd_loss_dict[classname].append(cd_loss.item())
chamfer_distance_list.append(cd_loss.item())
msg="processing %s ,class %s, cd loss: %f,mean cd_loss: %f, fscore: %f, mean fscore: %f" %(
result_file,classname,cd_loss.item(),np.mean(np.array(chamfer_distance_list)),fst,np.mean(np.array(fscore_list)))
print(msg)
with open(log_txt,'a') as f:
f.write(msg+"\n")
mean_chamfer_distance=np.mean(np.array(chamfer_distance_list))
msg="mean chamfer distance is %f"%(mean_chamfer_distance)
print(msg)
with open(log_txt, 'a') as f:
f.write(msg + "\n")
for key in cd_loss_dict:
cd_loss_dict[key]=np.mean(np.array(cd_loss_dict[key]))
for key in fst_dict:
fst_dict[key]=np.mean(np.array(fst_dict[key]))
for key in cd_loss_dict:
msg="cd loss of category %s is %f"%(key,cd_loss_dict[key])
print(msg)
with open(log_txt, 'a') as f:
f.write(msg + "\n")
for key in fst_dict:
msg="fscore of category %s is %f"%(key,fst_dict[key])
print(msg)
with open(log_txt,'a') as f:
f.write(msg+"\n")