Skip to content

S
stacked_hourglasses
Commit 86b64e4e by baihe
Options

updated

parent fa59fd22
Showing with 44 additions and 34 deletions
.gitignore
......@@ -11,5 +11,5 @@ _ext
tmp
*.o*
*~
.idea
.idea/
.DS_Store
\ No newline at end of file
.idea/pytorch_stacked_hourglass-master.iml
......@@ -14,7 +14,7 @@
<component name="PyNamespacePackagesService">
<option name="namespacePackageFolders">
<list>
<option value="$MODULE_DIR$/models" />
<option value="$MODULE_DIR$/mymodels" />
</list>
</option>
</component>
......
DetectorLoader.py
......@@ -22,8 +22,8 @@ class TinyYOLOv3_onecls(object):
"""
def __init__(self,
input_size=416,
config_file='models/yolo-tiny-onecls/yolov3-tiny-onecls.cfg',
weight_file='models/yolo-tiny-onecls/best-model.pth',
config_file='mymodels/yolo-tiny-onecls/yolov3-tiny-onecls.cfg',
weight_file='mymodels/yolo-tiny-onecls/best-model.pth',
nms=0.2,
conf_thres=0.45,
device='cuda'):
......
data/MPII/dp.py
......@@ -4,7 +4,7 @@ import os
import torch
import numpy as np
import torch.utils.data
import utils.img
import myutils.img
class GenerateHeatmap():
def __init__(self, output_res, num_parts):
......@@ -64,10 +64,10 @@ class Dataset(torch.utils.data.Dataset):
s = ds.get_scale(idx)
normalize = ds.get_normalized(idx)
cropped = utils.img.crop(orig_img, c, s, (self.input_res, self.input_res))
cropped = myutils.img.crop(orig_img, c, s, (self.input_res, self.input_res))
for i in range(np.shape(orig_keypoints)[1]):
if orig_keypoints[0,i,0] > 0:
orig_keypoints[0,i,:2] = utils.img.transform(orig_keypoints[0,i,:2], c, s, (self.input_res, self.input_res))
orig_keypoints[0,i,:2] = myutils.img.transform(orig_keypoints[0,i,:2], c, s, (self.input_res, self.input_res))
keypoints = np.copy(orig_keypoints)
## augmentation -- to be done to cropped image
......@@ -81,11 +81,11 @@ class Dataset(torch.utils.data.Dataset):
aug_scale = np.random.random() * (1.25 - 0.75) + 0.75
scale *= aug_scale
mat_mask = utils.img.get_transform(center, scale, (self.output_res, self.output_res), aug_rot)[:2]
mat_mask = myutils.img.get_transform(center, scale, (self.output_res, self.output_res), aug_rot)[:2]
mat = utils.img.get_transform(center, scale, (self.input_res, self.input_res), aug_rot)[:2]
mat = myutils.img.get_transform(center, scale, (self.input_res, self.input_res), aug_rot)[:2]
inp = cv2.warpAffine(cropped, mat, (self.input_res, self.input_res)).astype(np.float32)/255
keypoints[:,:,0:2] = utils.img.kpt_affine(keypoints[:,:,0:2], mat_mask)
keypoints[:,:,0:2] = myutils.img.kpt_affine(keypoints[:,:,0:2], mat_mask)
if np.random.randint(2) == 0:
inp = self.preprocess(inp)
inp = inp[:, ::-1]
......
handler.py
......@@ -2,9 +2,9 @@ import cv2
import torch
import data.MPII.ref as ds
import utils.img
from utils.group import HeatmapParser
from utils.posture import *
import myutils.img
from myutils.group import HeatmapParser
from myutils.posture import *
# 歪头的斜率阈值
CROOKED_HEAD_THRE=8
......@@ -105,13 +105,13 @@ def post_process(det, mat_, trainval, c=None, s=None, resolution=None):
cropped_preds = parser.parse(np.float32([det]))[0]
if len(cropped_preds) > 0:
cropped_preds[:, :, :2] = utils.img.kpt_affine(cropped_preds[:, :, :2] * 4, mat) # size 1x16x3
cropped_preds[:, :, :2] = myutils.img.kpt_affine(cropped_preds[:, :, :2] * 4, mat) # size 1x16x3
preds = np.copy(cropped_preds)
##for inverting predictions from input res on cropped to original image
if trainval != 'cropped':
for j in range(preds.shape[1]):
preds[0, j, :2] = utils.img.transform(preds[0, j, :2], c, s, resolution, invert=1)
preds[0, j, :2] = myutils.img.transform(preds[0, j, :2], c, s, resolution, invert=1)
return preds
......@@ -130,7 +130,7 @@ def inference(img, func, config, c, s):
res = (config['train']['input_res'], config['train']['input_res'])
# [[1,0,0],[0,1,0]]
mat_ = utils.img.get_transform(center, scale, res)[:2]
mat_ = myutils.img.get_transform(center, scale, res)[:2]
inp = img / 255
def array2dict(tmp):
......@@ -190,17 +190,26 @@ if __name__ == '__main__':
input_res = 256
orig_img = cv2.imread(image_path)
orig_img_reverse = cv2.imread(image_path)[:,:,::-1]
shape = orig_img_reverse.shape[0:2]
shape = orig_img_reverse.shape[0:2]
# 这里需要使用yolo检测图片中的人体矩形区域，返回矩形的中心和矩形较长边的边长/200
# 这里要起一个yolo5的服务。因为yolo5的模型的相对路径已经定死了，无法直接引入到项目中。除非通过命令行的方式
import requests
DETECTION_URL = "http://localhost:5000/v1/object-detection/yolov5s"
image_data = open(image_path, "rb").read()
response = requests.post(DETECTION_URL, files={"image": image_data}).json()
# import requests
# DETECTION_URL = "http://localhost:5000/v1/object-detection/yolov5s"
# image_data = open(image_path, "rb").read()
# response = requests.post(DETECTION_URL, files={"image": image_data}).json()
# for box in response:
# if box['class']!=0:
# continue
# else:
# c = [(box['xmax']+box['xmin'])/2,(box['ymax']+box['ymin'])/2]
# s = max((box['xmax']-box['xmin'])/200,(box['ymax']-box['ymin'])/200)+1
# Model
model = torch.hub.load('ultralytics/yolov5', 'yolov5s')
response = model(image_path).pandas().xyxy[0].to_dict(orient='records')
for box in response:
if box['class']!=0:
continue
......@@ -211,7 +220,7 @@ if __name__ == '__main__':
# # 矩形中心坐标点
# # 缩放比例 实验结果：720*1280的图片,人物占据大部分的情况下，7比较好。
# s = 7
im = utils.img.crop(orig_img_reverse,c,s,(input_res,input_res))
im = myutils.img.crop(orig_img_reverse,c,s,(input_res,input_res))
pred = do(im,c,s)
......
models/posenet.py → mymodels/posenet.py
import torch
from torch import nn
from models.layers import Conv, Hourglass, Pool, Residual
from mymodels.layers import Conv, Hourglass, Pool, Residual
from task.loss import HeatmapLoss
class UnFlatten(nn.Module):
......
task/pose.py
......@@ -7,11 +7,11 @@ import numpy as np
from torch import nn
import os
from torch.nn import DataParallel
from utils.misc import make_input, make_output, importNet
from myutils.misc import make_input, make_output, importNet
__config__ = {
'data_provider': 'data.MPII.dp',
'network': 'models.posenet.PoseNet',
'network': 'mymodels.posenet.PoseNet',
'inference': {
'nstack': 8,
'inp_dim': 256,
......@@ -80,7 +80,7 @@ def make_network(configs):
## creating new posenet
print(configs['network'])
configs['network'] = "models.posenet.PoseNet"
configs['network'] = "mymodels.posenet.PoseNet"
PoseNet = importNet(configs['network'])
poseNet = PoseNet(**config)
# forward_net = DataParallel(poseNet.cuda())
......
test.py
......@@ -6,8 +6,8 @@ import numpy as np
import h5py
import copy
from utils.group import HeatmapParser
import utils.img
from myutils.group import HeatmapParser
import myutils.img
import data.MPII.ref as ds
parser = HeatmapParser()
......@@ -18,13 +18,13 @@ def post_process(det, mat_, trainval, c=None, s=None, resolution=None):
cropped_preds = parser.parse(np.float32([det]))[0]
if len(cropped_preds) > 0:
cropped_preds[:,:,:2] = utils.img.kpt_affine(cropped_preds[:,:,:2] * 4, mat) #size 1x16x3
cropped_preds[:,:,:2] = myutils.img.kpt_affine(cropped_preds[:,:,:2] * 4, mat) #size 1x16x3
preds = np.copy(cropped_preds)
##for inverting predictions from input res on cropped to original image
if trainval != 'cropped':
for j in range(preds.shape[1]):
preds[0,j,:2] = utils.img.transform(preds[0,j,:2], c, s, resolution, invert=1)
preds[0,j,:2] = myutils.img.transform(preds[0,j,:2], c, s, resolution, invert=1)
return preds
def inference(img, func, config, c, s):
......@@ -38,7 +38,7 @@ def inference(img, func, config, c, s):
scale = max(height, width)/200
res = (config['train']['input_res'], config['train']['input_res'])
mat_ = utils.img.get_transform(center, scale, res)[:2]
mat_ = myutils.img.get_transform(center, scale, res)[:2]
inp = img/255
def array2dict(tmp):
......@@ -158,7 +158,7 @@ def get_img(config, num_eval=2958, num_train=300):
orig_img = cv2.imread(path_t)[:,:,::-1]
c = train_f['center'][i]
s = train_f['scale'][i]
im = utils.img.crop(orig_img, c, s, (input_res, input_res))
im = myutils.img.crop(orig_img, c, s, (input_res, input_res))
## kp
kp = train_f['part'][i]
......@@ -182,7 +182,7 @@ def get_img(config, num_eval=2958, num_train=300):
orig_img = cv2.imread(path_t)[:,:,::-1]
c = val_f['center'][i]
s = val_f['scale'][i]
im = utils.img.crop(orig_img, c, s, (input_res, input_res))
im = myutils.img.crop(orig_img, c, s, (input_res, input_res))
## kp
kp = val_f['part'][i]
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment