-
Notifications
You must be signed in to change notification settings - Fork 0
/
image.py
241 lines (192 loc) · 10.4 KB
/
image.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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
import os
import cv2
import json
import math
import time
import torch
import shutil
import imageio
import torchvision
import numpy as np
import torch.nn as nn
import torchvision.transforms as tfs
from scipy.spatial.transform import Rotation
from PIL import Image
from pathlib import Path
from tqdm import tqdm
from src.config import get_config
from src.agent import get_agent
from src.fisher.fisher_utils import batch_torch_A_to_R
from src.utils import compute_euler_angles_from_rotation_matrices
from src.utils import limit_angle
from src.vis_plot import convert_euler_bbox_to_6dof
from src.vis_plot import convert_rotmat_bbox_to_6dof
from src.renderer import Renderer
renderer = Renderer(
vertices_path="pose_references/vertices_trans.npy",
triangles_path="pose_references/triangles.npy"
)
###############################################################################
# BPJDet utils (the body-head model trained on CrowdHuman and using YOLOv5-L)
###############################################################################
from utils.torch_utils import select_device, time_sync
from utils.general import check_img_size, scale_coords, non_max_suppression
from utils.datasets import LoadImages
from models.experimental import attempt_load
colors_list = [
# [255, 0, 0], [255, 127, 0], [255, 255, 0], [127, 255, 0], [0, 255, 0], [0, 255, 127],
# [0, 255, 255], [0, 127, 255], [0, 0, 255], [127, 0, 255], [255, 0, 255], [255, 0, 127],
[255, 127, 0], [127, 255, 0], [0, 255, 127], [0, 127, 255], [127, 0, 255], [255, 0, 127],
[255, 255, 255],
[127, 0, 127], [0, 127, 127], [127, 127, 0], [127, 0, 0], [127, 0, 0], [0, 127, 0],
[127, 127, 127],
[255, 0, 255], [0, 255, 255], [255, 255, 0], [0, 0, 255], [255, 0, 0], [0, 255, 0],
[0, 0, 0],
[255, 127, 255], [127, 255, 255], [255, 255, 127], [127, 127, 255], [255, 127, 127], [255, 127, 127],
] # 27 colors
def cal_inside_iou(bigBox, smallBox): # body_box, part_box
# calculate small rectangle inside big box ratio, calSmallBoxInsideRatio
[Ax0, Ay0, Ax1, Ay1] = bigBox[0:4]
[Bx0, By0, Bx1, By1] = smallBox[0:4]
W = min(Ax1, Bx1) - max(Ax0, Bx0)
H = min(Ay1, By1) - max(Ay0, By0)
if W <= 0 or H <= 0:
return 0
else:
areaA = (Ax1 - Ax0)*(Ay1 - Ay0)
areaB = (Bx1 - Bx0)*(By1 - By0)
crossArea = W * H
# return crossArea/(areaA + areaB - crossArea)
return crossArea/areaB # range [0, 1]
def post_process_batch(imgs, paths, shapes, body_dets, part_dets, num_offsets, match_iou_thres):
batch_bboxes, batch_points, batch_scores = [], [], []
# process each image in batch
for si, (bdet, pdet) in enumerate(zip(body_dets, part_dets)):
nbody, npart = bdet.shape[0], pdet.shape[0]
if nbody: # one batch
path, shape = Path(paths[si]) if len(paths) else '', shapes[si][0]
scores = bdet[:, 4].cpu().numpy() # body detection score
bboxes = scale_coords(imgs[si].shape[1:], bdet[:, :4], shape).round().cpu().numpy()
points = scale_coords(imgs[si].shape[1:], bdet[:, -num_offsets:], shape).cpu().numpy()
points = points.reshape((nbody, -1, 2))
# points = np.concatenate((points, np.zeros((nbody, points.shape[1], 1))), axis=-1) # n*c*2 --> n*c*3
points = np.concatenate((points, np.zeros((nbody, points.shape[1], 5))), axis=-1) # n*c*2 --> n*c*7
# batch_parts_dict[str(img_id)] = []
if npart:
pdet[:, :4] = scale_coords(imgs[si].shape[1:], pdet[:, :4].clone(), shape)
pdet_slim = pdet[:, :6].cpu().numpy()
left_pdet = []
matched_part_ids = [-1 for i in range(points.shape[0])] # points shape is n*c*7, add in 2022-12-09
for id, (x1, y1, x2, y2, conf, cls) in enumerate(pdet_slim):
p_xc, p_yc = np.mean((x1, x2)), np.mean((y1, y2)) # the body-part's part bbox center point
part_pts = points[:, int(cls - 1)]
dist = np.linalg.norm(part_pts[:, :2] - np.array([[p_xc, p_yc]]), axis=-1)
pt_match = np.argmin(dist)
'''association alg version 4.0'''
tmp_iou = cal_inside_iou(bboxes[pt_match], [x1, y1, x2, y2]) # add in 2022-12-11, body-part must inside the body
if conf > part_pts[pt_match][2] and tmp_iou > match_iou_thres: # add in 2022-12-09, we fetch the part bbox with highest conf
part_pts[pt_match] = [p_xc, p_yc, conf, x1, y1, x2, y2] # update points[:, int(cls - 1), 7]
matched_part_ids[pt_match] = id # only for dataset BodyHands without left/right label of hands
batch_bboxes.extend(bboxes)
batch_points.extend(points)
batch_scores.extend(scores)
# else:
# print("This image has no object detected!")
return batch_bboxes, batch_points, batch_scores
###############################################################################
if __name__ == '__main__':
alpha_value = 1.0
edges_scale = -0.05
config = get_config('test')
agent = get_agent(config)
agent.load_ckpt(config.ckpt)
"""Step 0: get single image / multiple images"""
# folder_name, imgs_name = "test_DetHPE", "SCUTHead" # classrooms, meeting rooms
# folder_name, imgs_name = "test_DetHPE", "CroHD" # crossing, stations
folder_name, imgs_name = "test_DetHPE", "CrowdHuman" # crowded persons
image_path = f"./{folder_name}/{imgs_name}"
"""Step 1: load model weights"""
weights_path = "weights/ch_head_l_1536_e150_best_mMR.pt"
conf_thres = 0.5
iou_thres = 0.75
match_iou_thres = 0.6
scales = [1]
imgsz = 1536
num_offsets = 2
body_bbox, head_bbox = False, True
device = 'cuda:0'
print('Using device: {}'.format(device))
BPJDet_model = attempt_load(weights_path, map_location=device) # load FP32 model
stride = int(BPJDet_model.stride.max()) # model stride
imgsz = check_img_size(imgsz, s=stride) # check image size
dataset = LoadImages(image_path, img_size=imgsz, stride=stride, auto=True)
dataset_iter = iter(dataset)
"""Step 2: processing images"""
print(image_path, len(dataset))
for index in range(len(dataset)):
(single_path, img, im0, _) = next(dataset_iter)
if '_res' in single_path or '_vis' in single_path:
continue
print(index, single_path, "\n")
img = torch.from_numpy(img).to(device)
img = img / 255.0 # 0 - 255 to 0.0 - 1.0
if len(img.shape) == 3:
img = img[None] # expand for batch dim
out_ori = BPJDet_model(img, augment=True, scales=scales)[0]
body_dets = non_max_suppression(out_ori, conf_thres, iou_thres,
classes=[0], num_offsets=num_offsets)
part_dets = non_max_suppression(out_ori, conf_thres, iou_thres,
classes=list(range(1, 1 + num_offsets//2)), num_offsets=num_offsets)
# Post-processing of body and part detections
bboxes, points, scores = post_process_batch(
img, [], [[im0.shape[:2]]], body_dets, part_dets, num_offsets, match_iou_thres)
line_thick = max(im0.shape[:2]) // 1000 + 3
im0_ori = im0.copy()
img_h, img_w, img_c = im0_ori.shape
head_bbox_list = []
for ind, (bbox, point, score) in enumerate(zip(bboxes, points, scores)):
f_score, f_bbox = point[0][2], point[0][3:] # bbox format [x1, y1, x2, y2]
if f_score != 0: # for the body-head pair, we must have a detected head
color = colors_list[ind%len(colors_list)]
head_bbox_list.append(f_bbox)
if body_bbox:
[x1, y1, x2, y2] = bbox
cv2.rectangle(im0, (int(x1), int(y1)), (int(x2), int(y2)), color, thickness=line_thick)
if head_bbox:
[px1, py1, px2, py2] = f_bbox
cv2.rectangle(im0, (int(px1), int(py1)), (int(px2), int(py2)), color, thickness=line_thick)
for [px1, py1, px2, py2] in head_bbox_list:
pcx, pcy = (px1 + px2)/2.0, (py1 + py2)/2.0
head_size = max(px2 - px1, py2 - py1)
new_px1 = max(0, int(pcx - (0.5 - edges_scale) * head_size))
new_px2 = min(img_w-1, int(pcx + (0.5 - edges_scale) * head_size))
new_py1 = max(0, int(pcy - (0.5 - edges_scale) * head_size))
new_py2 = min(img_h-1, int(pcy + (0.5 - edges_scale) * head_size))
img_head_cv2 = im0_ori[new_py1:new_py2, new_px1:new_px2]
img_head_pil = img_head_cv2[:, :, ::-1] # Convert BGR to RGB
img_head_pil = Image.fromarray(img_head_pil)
img_input = img_head_pil.resize([224, 224])
img_tensor = tfs.ToTensor()(img_input)
img_tensor = tfs.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])(img_tensor)
with torch.no_grad():
agent.net.eval()
fisher_out = agent.net(img_tensor.reshape((1,3,224,224)).cuda())
pd_m = batch_torch_A_to_R(fisher_out)
rot_mat = pd_m.detach().cpu().numpy()[0]
rot_mat_2 = np.transpose(rot_mat)
angle = Rotation.from_matrix(rot_mat_2).as_euler("xyz", degrees=True)
roll, pitch, yaw = list(map(limit_angle, [angle[2], angle[0] - 180, angle[1]]))
head_bbox = [new_px1, new_py1, new_px2, new_py2]
global_intrinsics = np.array([[img_w+img_h, 0, img_w // 2], [0, img_w+img_h, img_h // 2], [0, 0, 1]])
global_pose = convert_rotmat_bbox_to_6dof(rot_mat, head_bbox, global_intrinsics)
trans_vertices = renderer.transform_vertices(im0, [global_pose])
im0 = renderer.render(im0, trans_vertices, alpha=alpha_value)
cv2.putText(im0, '{:2d} Heads'.format(len(head_bbox_list)), (5, 25),
cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 255, 255), thickness=2)
cv2.imwrite(single_path[:-4]+"_res_BPJDetSemiUHPE.jpg", im0)
'''
python image.py SSL1.0_r0.05_ce_tDyna0.75_b16_ema_RO_CO_CM_full/Sep20_195132/best \
--is_full_range --config settings/DAD3DHeads_COCOHead.yml --network resnet50 --gpu_ids 1
python image.py SSL1.0_r0.05_ce_tDyna0.75_b32_ema_RO_CO_CM_full/Sep30_130637/best \
--is_full_range --config settings/DAD3DHeads_COCOHead.yml --network repvgg --gpu_ids 1
'''