def __getitem__(self, idx):
if self.is_trian:
ele_anno = self.anno[self.train_list[idx]]
else:
ele_anno = self.anno[self.valid_list[idx]]
res_img = [256, 256]
res_heatmap = [64, 64]
path_img_folder = 'mpii_human_pose_v1/images'
path_img = os.path.join(path_img_folder, ele_anno['img_paths'])
img_origin = imgutils.load_image(path_img)
img_crop, ary_pts_crop, c_crop = imgutils.crop(
img_origin,
ele_anno,
use_randscale=self.use_scale,
use_randflipLR=self.use_flip,
use_randcolor=self.use_rand_color)
img_out, pts_out, c_out = imgutils.change_resolu(
img_crop, ary_pts_crop, c_crop, res_heatmap)
train_img = skimage.transform.resize(img_crop, tuple(res_img))
train_heatmap = imgutils.generate_heatmaps(img_out,
pts_out,
sigma_valu=2)
train_pts = pts_out[:, :2].astype(np.int32)
# (H,W,C) -> (C,H,W)
train_img = np.transpose(train_img, (2, 0, 1))
train_heatmap = np.transpose(train_heatmap, (2, 0, 1))
return train_img, train_heatmap, train_pts
def build_input(self, imgpaths):
"""
build input data for model input tensor
:param imgpaths: images path array
:return: 3 channel image array
"""
images = np.array([imgutils.load_image(img_path) for img_path in imgpaths])
return [img.reshape(224, 224, 3) for img in images]
from utils import imgutilsTF
from utils import Dataset_Make_Utils as dsmutils
from utils import modelutils
from models.hourglass_new import Residual, HourGlass, HourGlassNet, hg
from models.hourglass import hg as hg_torch
from losses.jointMSE import mse_joint as jMSE
from datasets.dataset_torch import DatasetTorch
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
#========================================================== TEST BY NEW IMGS ==========================================================
# Load test img
num_img = 6
path_testimg = 'test_imgs/' + str(num_img) + '.jpg'
img_np = imgutils.load_image(path_testimg)
print('img_np.SAPE', img_np.shape)
# Resize to (256,256,3)
img_np = skimage.transform.resize(img_np, [256, 256])
img_np_copy = img_np
print('img_np.SAPE', img_np.shape)
# plt.figure(1)
# plt.imshow(img_np_copy)
# plt.show()
# # ================================== Load the model ==================================
# num_subset = 'allEPOCH2'
# net_hg_torch = hg_torch(num_stacks=8, num_blocks=1, num_classes=16)
# path_model_torch_load = 'models/modelparams_hg_torch_' + str(num_subset) + '.pkl'
train_heatmaps = np.zeros((len_idx, res_heatmap[0], res_heatmap[1], 16)) #(N,H_im,W_im,3)
path_img_folder = path_dir+'/mpii_human_pose_v1/images'
num_subset = '4000(3)' # 0 means all
j = 0
for i, idx_train in enumerate(train_list):
if i >= idx_begin and i <= idx_end:
j += 1
print(anno[idx_train])
ele_anno = anno[idx_train]
path_img = os.path.join(path_img_folder, ele_anno['img_paths'])
print(path_img)
img_origin = imgutils.load_image(path_img)
# plt.figure(1)
# plt.imshow(img_origin)
# plt.show()
img_crop, ary_pts_crop, c_crop = imgutils.crop_norescale(img_origin, ele_anno, use_randscale=False)
img_out, pts_out, c_out = imgutils.change_resolu(img_crop, ary_pts_crop, c_crop, res_heatmap)
heatmaps = imgutils.generate_heatmaps(img_out, pts_out, sigma_valu=2)
# plt.figure(1)
# plt.imshow(img_out)
# plt.show()
train_heatmaps[j - 1, ...] = heatmaps
# plt.figure(1)
# plt.imshow(train_heatmaps[j-1, :, :, 0])
# plt.show()
# imgutils.show_stack_joints(img_out, pts_out, c_out)
# print('heatmaps.SHAPE', train_heatmaps[j - 1, ...].shape)
# # Show points on image
# imgutils.show_stack_joints(img_lowres, ary_pts, c)
#=======================================================================================
# Get json annotation
with open('mpii_annotations.json') as anno_file:
anno = json.load(anno_file)
for ele in anno:
if ele['img_paths'] == '000678817.jpg': #000033016, 000678817, 001834988
ele_copy = ele
print('FOUNDED')
print(ele)
c_copy = ele_copy['objpos']
pts_copy = ele_copy['joint_self']
img_origin = imgutils.load_image('mpii_human_pose_v1/images/000678817.jpg')
img_copy = img_origin
# img_crop, ary_pts_crop, c_crop = imgutils.crop(img_origin, ele_copy)
img_crop, ary_pts_crop, c_crop = imgutils.crop_norescale(img_origin, ele_copy)
img_out, pts_out, c_out = imgutils.change_resolu(img_crop, ary_pts_crop,
c_crop, (256, 256))
heatmaps = imgutils.generate_heatmaps(img_out, pts_out, sigma_valu=7)
# img_out = imgutils.random_rotate(img_out)
# heatmaps = imgutils.random_rotate(heatmaps[:,:,:])
# img_out = imgutilsTF.random_flip_tf(img_out)
# heatmaps = imgutilsTF.random_flip_tf(heatmaps[:,:,:])
# Show image and heatmaps
imgutils.show_stack_joints(img_out, pts_out, c_out)
