in eft/datasets/video_dataset.py [0:0]
def __getitem__(self, index):
bLoadImage = self.bLoadImage
# print("Load: {}".format(index) )
if g_debugMode: ############### DEBUG MODE ###############
import os
if os.path.exists("exemplarTestTarget.txt"): #Read target index from this file
with open("exemplarTestTarget.txt","r") as f:
targetIdx = f.readlines()[0]
print("targetImgFrame: {}".format(targetIdx))
self.options.exemplar_targetIdx = targetIdx #Save this. If a new one model should be reinitiated in base_trainer.py
else:
targetIdx = '001'
for i in range(len(self.imgname)):
if targetIdx in self.imgname[i]:
index =i
break
print("DEBUG: selected index {}".format(index))
# for i in range(len(self.imgname)):
# # if 'COCO_train2014_000000000322' in self.imgname[i]:
# if '00_15_00030840' in self.imgname[i]:
# index =i
# break
# print("DEBUG: selected index {}".format(index))
#Good panoptic sample: 171204_pose6/00003110/00_00_00003110.jpg
item = {}
if self.bEnforceUpperOnly: #Always upper only
if self.scale_upperbody[index]>0:
scale = self.scale_upperbody[index].copy()
center = self.center_upperbody[index].copy()
else: #except no valid cropping
scale = self.scale[index].copy()
center = self.center[index].copy()
elif self.is_train and (self.options is not None and self.options.bUpperBodyTest):
randomNum =np.random.uniform()
# print(randomNum)
if self.scale_upperbody[index]>0 and randomNum <= self.options.upperBodyTest_prob:
if self.options.bFacePartTest:
randomNumAgain =np.random.uniform()
if randomNumAgain<0.5 and self.scale_facePart[index]>0:
# print("facepart")
scale = self.scale_facePart[index].copy()
center = self.center_facePart[index].copy()
else:
# print("upperbody")
scale = self.scale_upperbody[index].copy()
center = self.center_upperbody[index].copy()
else:
scale = self.scale_upperbody[index].copy()
center = self.center_upperbody[index].copy()
else:
scale = self.scale[index].copy()
center = self.center[index].copy()
else:
scale = self.scale[index].copy()
center = self.center[index].copy()
# Get augmentation parameters
if (self.options is not None and self.options.bExemplarMode) or g_debugMode: #Debug: No augmentation#g_debugMode: ############### DEBUG MODE ###############
flip = 0 # flipping
pn = np.ones(3) # per channel pixel-noise
rot = 0 # rotation
sc = 1
else:
flip, pn, rot, sc = self.augm_params()
# Load image
imgname = join(self.img_dir, self.imgname[index])
# print("dbName: {} | imgname: {}".format(self.datasetName, imgname))
if bLoadImage:
try:
img = cv2.imread(imgname)[:,:,::-1].copy().astype(np.float32)
except TypeError:
print("Error: cannnot find image from: {}".format(imgname) )
orig_shape = np.array(img.shape)[:2]
else:
orig_shape =0
# Get SMPL parameters, if available
if self.has_smpl[index]:
pose = self.pose[index].copy()
betas = self.betas[index].copy()
else:
pose = np.zeros(72)
betas = np.zeros(10)
# Process image
if bLoadImage:
try:
img = self.rgb_processing(img, center, sc*scale, rot, flip, pn)
except:
# print("Warning: image processing has been failed: {}".format(imgname))
print("Warning: image processing has been failed: {}: sc{}, rot{}, flipt{},pn{}".format(imgname,sc,rot,flip,pn))
img = np.zeros((3,224,224), dtype=np.float32) #generate blank image
if img is None:
# print("Warning: image processing has been failed: {}".format(imgname))
print("Warning: image processing has been failed: {}: sc{}, rot{}, flipt{},pn{}".format(imgname,sc,rot,flip,pn))
img = np.zeros((3,224,224), dtype=np.float32) #generate blank image
img = torch.from_numpy(img).float()
item['img'] = self.normalize_img(img)
else:
item['img'] =''
# Store image before normalization to use it in visualization
item['pose'] = torch.from_numpy(self.pose_processing(pose, rot, flip)).float()
item['betas'] = torch.from_numpy(betas).float()
item['imgname'] = imgname
if self.annotIds is not None:
item['annotId'] = self.annotIds[index]
else:
item['annotId'] = -1 #Garbage
# print("Debug: {}".format(item['annotId']))
# Get 3D pose, if available
if self.has_pose_3d:
S = self.pose_3d[index].copy()
item['pose_3d'] = torch.from_numpy(self.j3d_processing(S, rot, flip)).float()
else:
item['pose_3d'] = torch.zeros(24,4, dtype=torch.float32)
#Processing Hands
if self.options.bUseHand3D:
S = np.ones( (self.rhand_3d[index].shape[0], self.rhand_3d[index].shape[1]+1) )
S[:,:-1] = self.rhand_3d[index].copy() #
item['rhand_3d'] = torch.from_numpy(self.j3d_processing(S, rot, flip)).float()
S = np.ones( (self.lhand_3d[index].shape[0], self.lhand_3d[index].shape[1]+1) )
S[:,:-1] = self.lhand_3d[index].copy()
item['lhand_3d'] = torch.from_numpy(self.j3d_processing(S, rot, flip)).float()
# else:
# item['pose_3d'] = torch.zeros(24,4, dtype=torch.float32)
# Get 2D keypoints and apply augmentation transforms
keypoints = self.keypoints[index].copy()
#Disable specifically foot, if too close to boundarys
imgHeight = orig_shape[0]
if abs(keypoints[25+0,1] - imgHeight)<10 and keypoints[10,2]<0.1: #Right Foot. within 10 pix from the boundary
keypoints[25+0,2] = 0 #Disable
if abs(keypoints[25+5,1] - imgHeight)<10 and keypoints[13,2]<0.1: #Left Foot.
keypoints[25+5,2] =0 #Disable
item['keypoints_original'] = self.keypoints[index].copy()
item['keypoints'] = torch.from_numpy(self.j2d_processing(keypoints, center, sc*scale, rot, flip)).float()
item['has_smpl'] = self.has_smpl[index]
item['has_pose_3d'] = self.has_pose_3d
item['scale'] = float(sc * scale)
item['center'] = center.astype(np.float32)
item['orig_shape'] = orig_shape
item['is_flipped'] = flip
item['rot_angle'] = np.float32(rot)
item['gender'] = self.gender[index]
item['sample_index'] = index
item['dataset_name'] = self.datasetName
#for Panoptic DB.... to check existence...bad code.
# parentIdx = 100* (int(index /100.0) + 1)
# item['pkl_save_name'] = join(self.img_dir,self.imgname[parentIdx])
try:
item['maskname'] = self.maskname[index]
except AttributeError:
item['maskname'] = ''
try:
item['partname'] = self.partname[index]
except AttributeError:
item['partname'] = ''
return item