def visEFT_singleSubject(inputDir, imDir, smplModelDir, bUseSMPLX):
if bUseSMPLX:
smpl = SMPLX(smplModelDir, batch_size=1, create_transl=False)
else:
smpl = SMPL(smplModelDir, batch_size=1, create_transl=False)
fileList = listdir(inputDir) #Check all fitting files
print(">> Found {} files in the fitting folder {}".format(len(fileList), inputDir))
totalCnt =0
erroneousCnt =0
# fileList =['00_00_00008422_0.pkl', '00_00_00008422_1731.pkl', '00_00_00008422_3462.pkl'] #debug
for f in sorted(fileList):
#Load
fileFullPath = join(inputDir, f)
with open(fileFullPath,'rb') as f:
dataDict = pickle.load(f)
print(f"Loaded :{fileFullPath}")
if 'imageName' in dataDict.keys(): #If this pkl has only one instance. Made this to hand panoptic output where pkl has multi instances
dataDict = {0:dataDict}
for jj, k in enumerate(dataDict):
if jj%50 !=0:
continue
data = dataDict[k]
# print(data['subjectId'])
# continue
if 'smpltype' in data:
if (data['smpltype'] =='smpl' and bUseSMPLX) or (data['smpltype'] =='smplx' and bUseSMPLX==False):
print("SMPL type mismatch error")
assert False
imgFullPathOri = data['imageName'][0]
imgFullPath = os.path.join(imDir, os.path.basename(imgFullPathOri))
data['subjectId'] =0 #TODO debug
fileName = "{}_{}".format(data['subjectId'], os.path.basename(imgFullPathOri)[:-4])
if args.bRenderToFiles and os.path.exists(os.path.join(render_dirName, fileName+".jpg")):
continue
if True: #Additional path checking, if not valid
if os.path.exists(imgFullPath) == False:
imgFullPath =getpath_level(imDir, imgFullPathOri ,1)
if os.path.exists(imgFullPath) == False:
imgFullPath =getpath_level(imDir, imgFullPathOri,2)
if os.path.exists(imgFullPath) == False:
imgFullPath =getpath_level(imDir, imgFullPathOri, 3 )
scale = data['scale'][0]
center = data['center'][0]
# print(data['annotId'])
ours_betas = torch.from_numpy(data['pred_shape'])
ours_pose_rotmat = torch.from_numpy(data['pred_pose_rotmat'])
# spin_betas = torch.from_numpy(data['opt_beta'])
#Compute 2D reprojection error
# if not (data['loss_keypoints_2d']<0.0001 or data['loss_keypoints_2d']>0.001 :
# continue
maxBeta = abs(torch.max( abs(ours_betas)).item())
if data['loss_keypoints_2d']>0.0005 or maxBeta>3:
erroneousCnt +=1
print(">>> loss2d: {}, maxBeta: {}".format( data['loss_keypoints_2d'],maxBeta) )
# spin_pose = torch.from_numpy(data['opt_pose'])
pred_camera_vis = data['pred_camera']
if os.path.exists(imgFullPath) == False:
print(imgFullPath)
assert os.path.exists(imgFullPath)
rawImg = cv2.imread(imgFullPath)
print(imgFullPath)
croppedImg, boxScale_o2n, bboxTopLeft = crop_bboxInfo(rawImg, center, scale, (BBOX_IMG_RES, BBOX_IMG_RES) )
#Visualize 2D image
if args.bRenderToFiles ==False:
viewer2D.ImShow(rawImg, name='rawImg', waitTime=10) #You should press any key
viewer2D.ImShow(croppedImg, name='croppedImg', waitTime=10)
ours_output = smpl(betas=ours_betas, body_pose=ours_pose_rotmat[:,1:], global_orient=ours_pose_rotmat[:,0].unsqueeze(1), pose2rot=False )
ours_vertices = ours_output.vertices.detach().cpu().numpy()
ours_joints_3d = ours_output.joints.detach().cpu().numpy()
#Visualize 3D mesh and 3D skeleton in BBox Space
if True:
b =0
camParam_scale = pred_camera_vis[b,0]
camParam_trans = pred_camera_vis[b,1:]
############### Visualize Mesh ###############
pred_vert_vis = ours_vertices[b].copy()
pred_vert_vis = convert_smpl_to_bbox(pred_vert_vis, camParam_scale, camParam_trans)
pred_meshes = {'ver': pred_vert_vis, 'f': smpl.faces}
glViewer.setMeshData([pred_meshes], bComputeNormal= True)
################ Visualize Skeletons ###############
#Vis pred-SMPL joint
pred_joints_vis = ours_joints_3d[b,:,:3].copy() #(N,3)
pred_joints_vis = convert_smpl_to_bbox(pred_joints_vis, camParam_scale, camParam_trans)
glViewer.setSkeleton( [pred_joints_vis.ravel()[:,np.newaxis]])
################ Other 3D setup###############
glViewer.setBackgroundTexture(croppedImg)
glViewer.setWindowSize(croppedImg.shape[1]*args.windowscale, croppedImg.shape[0]*args.windowscale)
glViewer.SetOrthoCamera(True)
print("Press 'q' in the 3D window to go to the next sample")
glViewer.show(0)
#Visualize 3D mesh and 3D skeleton on original image space
if True:
b =0
camParam_scale = pred_camera_vis[b,0]
camParam_trans = pred_camera_vis[b,1:]
############### Visualize Mesh ###############
pred_vert_vis = ours_vertices[b].copy()
pred_vert_vis = convert_smpl_to_bbox(pred_vert_vis, camParam_scale, camParam_trans)
#From cropped space to original
pred_vert_vis = convert_bbox_to_oriIm(pred_vert_vis, boxScale_o2n, bboxTopLeft, rawImg.shape[1], rawImg.shape[0])
pred_meshes = {'ver': pred_vert_vis, 'f': smpl.faces}
glViewer.setMeshData([pred_meshes], bComputeNormal= True)
# ################ Visualize Skeletons ###############
#Vis pred-SMPL joint
pred_joints_vis = ours_joints_3d[b,:,:3].copy() #(N,3)
pred_joints_vis = convert_smpl_to_bbox(pred_joints_vis, camParam_scale, camParam_trans)
pred_joints_vis = convert_bbox_to_oriIm(pred_joints_vis, boxScale_o2n, bboxTopLeft, rawImg.shape[1], rawImg.shape[0])
glViewer.setSkeleton( [pred_joints_vis.ravel()[:,np.newaxis]])
glViewer.setBackgroundTexture(rawImg)
glViewer.setWindowSize(rawImg.shape[1]*args.magnifyFactor, rawImg.shape[0]*args.magnifyFactor)
glViewer.SetOrthoCamera(True)
print("Press 'q' in the 3D window to go to the next sample")
if args.bRenderToFiles: #Export rendered files
if os.path.exists(render_dirName) == False: #make a output folder if necessary
os.mkdir(render_dirName)
# subjId = data['subjectId'][22:24]
fileName = "{}_{}".format(data['subjectId'], os.path.basename(imgFullPathOri)[:-4])
# rawImg = cv2.putText(rawImg,data['subjectId'],(100,100), cv2.FONT_HERSHEY_PLAIN, 2, (255,255,0),2)
glViewer.render_on_image(render_dirName, fileName, rawImg)
print(f"Render to {fileName}")
def init_fn(self):
self.train_ds = MixedDataset(self.options,
ignore_3d=self.options.ignore_3d,
is_train=True)
self.model = hmr(config.SMPL_MEAN_PARAMS,
pretrained=True).to(self.device)
if self.options.bExemplarMode:
lr = 5e-5 * 0.2
else:
lr = self.options.lr
self.optimizer = torch.optim.Adam(
params=self.model.parameters(),
# lr=self.options.lr,
lr=lr,
weight_decay=0)
if self.options.bUseSMPLX: #SMPL-X model #No change is required for HMR training. SMPL-X ignores hand and other parts.
#SMPL uses 23 joints, while SMPL-X uses 21 joints, automatically ignoring the last two joints of SMPL
self.smpl = SMPLX(config.SMPL_MODEL_DIR,
batch_size=self.options.batch_size,
create_transl=False).to(self.device)
else: #Original SMPL
self.smpl = SMPL(config.SMPL_MODEL_DIR,
batch_size=self.options.batch_size,
create_transl=False).to(self.device)
# Per-vertex loss on the shape
self.criterion_shape = nn.L1Loss().to(self.device)
# Keypoint (2D and 3D) loss
# No reduction because confidence weighting needs to be applied
self.criterion_keypoints = nn.MSELoss(reduction='none').to(self.device)
# Loss for SMPL parameter regression
self.criterion_regr = nn.MSELoss().to(self.device)
self.models_dict = {'model': self.model}
self.optimizers_dict = {'optimizer': self.optimizer}
self.focal_length = constants.FOCAL_LENGTH
# Initialize SMPLify fitting module
self.smplify = SMPLify(step_size=1e-2,
batch_size=self.options.batch_size,
num_iters=self.options.num_smplify_iters,
focal_length=self.focal_length)
if self.options.pretrained_checkpoint is not None:
print(">>> Load Pretrained mode: {}".format(
self.options.pretrained_checkpoint))
self.load_pretrained(
checkpoint_file=self.options.pretrained_checkpoint)
self.backupModel()
#This should be called here after loading model
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
self.model = torch.nn.DataParallel(self.model) #Failed...
# Load dictionary of fits
self.fits_dict = FitsDict(self.options, self.train_ds)
# Create renderer
self.renderer = None # Renderer(focal_length=self.focal_length, img_res=self.options.img_res, faces=self.smpl.faces)
#debug
from torchvision.transforms import Normalize
self.de_normalize_img = Normalize(mean=[
-constants.IMG_NORM_MEAN[0] / constants.IMG_NORM_STD[0],
-constants.IMG_NORM_MEAN[1] / constants.IMG_NORM_STD[1],
-constants.IMG_NORM_MEAN[2] / constants.IMG_NORM_STD[2]
],
std=[
1 / constants.IMG_NORM_STD[0],
1 / constants.IMG_NORM_STD[1],
1 / constants.IMG_NORM_STD[2]
])
def visEFT_multiSubjects(inputDir, imDir, smplModelDir, bUseSMPLX = False):
if bUseSMPLX:
smpl = SMPLX(smplModelDir, batch_size=1, create_transl=False)
else:
smpl = SMPL(smplModelDir, batch_size=1, create_transl=False)
fileList = listdir(inputDir) #Check all fitting files
print(">> Found {} files in the fitting folder {}".format(len(fileList), inputDir))
totalCnt =0
erroneousCnt =0
#Merge sample from the same image
data_perimage ={}
for f in sorted(fileList):
if "_init" in f:
continue
#Load
imageName = f[:f.rfind('_')]
if imageName not in data_perimage.keys():
data_perimage[imageName] =[]
data_perimage[imageName].append(f)
for imgName in data_perimage:
eftFileNames = data_perimage[imgName]
meshData =[]
skelData =[]
for f in eftFileNames:
fileFullPath = join(inputDir, f)
with open(fileFullPath,'rb') as f:
data = pickle.load(f)
imgFullPathOri = data['imageName'][0]
imgFullPath = os.path.join(imDir, os.path.basename(imgFullPathOri))
if True: #Additional path checking, if not valid
if os.path.exists(imgFullPath) == False:
imgFullPath =getpath_level(imDir, imgFullPathOri ,1)
if os.path.exists(imgFullPath) == False:
imgFullPath =getpath_level(imDir, imgFullPathOri,2)
if os.path.exists(imgFullPath) == False:
imgFullPath =getpath_level(imDir, imgFullPathOri, 3 )
scale = data['scale'][0]
center = data['center'][0]
ours_betas = torch.from_numpy(data['pred_shape'])
ours_pose_rotmat = torch.from_numpy(data['pred_pose_rotmat'])
# spin_betas = torch.from_numpy(data['opt_beta'])
#Compute 2D reprojection error
# if not (data['loss_keypoints_2d']<0.0001 or data['loss_keypoints_2d']>0.001 :
# continue
maxBeta = abs(torch.max( abs(ours_betas)).item())
if data['loss_keypoints_2d']>0.0005 or maxBeta>3:
erroneousCnt +=1
print(">>> loss2d: {}, maxBeta: {}".format( data['loss_keypoints_2d'],maxBeta) )
# spin_pose = torch.from_numpy(data['opt_pose'])
pred_camera_vis = data['pred_camera']
assert os.path.exists(imgFullPath)
rawImg = cv2.imread(imgFullPath)
print(imgFullPath)
croppedImg, boxScale_o2n, bboxTopLeft = crop_bboxInfo(rawImg, center, scale, (constants.IMG_RES, constants.IMG_RES) )
#Visualize 2D image
if args.bRenderToFiles ==False:
viewer2D.ImShow(rawImg, name='rawImg', waitTime=10) #You should press any key
viewer2D.ImShow(croppedImg, name='croppedImg', waitTime=10)
if bUseSMPLX:
ours_output = smpl(betas=ours_betas, body_pose=ours_pose_rotmat[:,1:-2], global_orient=ours_pose_rotmat[:,0].unsqueeze(1), pose2rot=False )
# ours_output = smpl() #Default test
else:
ours_output = smpl(betas=ours_betas, body_pose=ours_pose_rotmat[:,1:], global_orient=ours_pose_rotmat[:,0].unsqueeze(1), pose2rot=False )
# ours_output = smpl() #Default test
ours_vertices = ours_output.vertices.detach().cpu().numpy()
ours_joints_3d = ours_output.joints.detach().cpu().numpy()
if False: #Debugging
# ours_vertices = ours_vertices - ours_joints_3d[0,12,:]
save_mesh_obj(ours_vertices[0], smpl.faces, 'test.obj')
#Visualize 3D mesh and 3D skeleton on original image space
if True:
b =0
camParam_scale = pred_camera_vis[b,0]
camParam_trans = pred_camera_vis[b,1:]
############### Visualize Mesh ###############
pred_vert_vis = ours_vertices[b].copy()
pred_vert_vis = convert_smpl_to_bbox(pred_vert_vis, camParam_scale, camParam_trans)
#From cropped space to original
pred_vert_vis = convert_bbox_to_oriIm(pred_vert_vis, boxScale_o2n, bboxTopLeft, rawImg.shape[1], rawImg.shape[0])
pred_meshes = {'ver': pred_vert_vis, 'f': smpl.faces}
# glViewer.setMeshData([pred_meshes], bComputeNormal= True)
# ################ Visualize Skeletons ###############
#Vis pred-SMPL joint
# pred_joints_vis = ours_joints_3d[b,-9:,:3].copy() #(N,3) #Debuggin
pred_joints_vis = ours_joints_3d[b,:,:3].copy() #(N,3)
pred_joints_vis = convert_smpl_to_bbox(pred_joints_vis, camParam_scale, camParam_trans)
pred_joints_vis = convert_bbox_to_oriIm(pred_joints_vis, boxScale_o2n, bboxTopLeft, rawImg.shape[1], rawImg.shape[0])
meshData.append(pred_meshes)
skelData.append(pred_joints_vis.ravel()[:,np.newaxis])
# glViewer.setSkeleton( [pred_joints_vis.ravel()[:,np.newaxis]])
glViewer.setBackgroundTexture(rawImg)
glViewer.setWindowSize(rawImg.shape[1]*args.magnifyFactor, rawImg.shape[0]*args.magnifyFactor)
glViewer.SetOrthoCamera(True)
# print("Press 'q' in the 3D window to go to the next sample")
# glViewer.show(0)
glViewer.setSkeleton(skelData)
glViewer.setMeshData(meshData, bComputeNormal= True)
if args.bRenderToFiles: #Export rendered files
if os.path.exists(render_dirName) == False: #make a output folder if necessary
os.mkdir(render_dirName)
fileName = imgFullPathOri[:-4].replace("/","_")
glViewer.render_on_image(render_dirName, fileName, rawImg)
print(f"render to {fileName}")
glViewer.show(args.displaytime)
def run_evaluation(model, dataset_name, dataset, result_file,
batch_size=32, img_res=224,
num_workers=32, shuffle=False, log_freq=50, bVerbose= True):
"""Run evaluation on the datasets and metrics we report in the paper. """
print(dataset_name)
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
# # Transfer model to the GPU
# model.to(device)
# Load SMPL model
global g_smpl_neutral, g_smpl_male, g_smpl_female
if g_smpl_neutral is None:
g_smpl_neutral = SMPL(config.SMPL_MODEL_DIR,
create_transl=False).to(device)
# g_smpl_neutral = SMPLX(config.SMPL_MODEL_DIR,
# create_transl=False).to(device)
g_smpl_male = SMPL(config.SMPL_MODEL_DIR,
gender='male',
create_transl=False).to(device)
g_smpl_female = SMPL(config.SMPL_MODEL_DIR,
gender='female',
create_transl=False).to(device)
smpl_neutral = g_smpl_neutral
smpl_male = g_smpl_male
smpl_female = g_smpl_female
else:
smpl_neutral = g_smpl_neutral
smpl_male = g_smpl_male
smpl_female = g_smpl_female
# renderer = PartRenderer()
# Regressor for H36m joints
J_regressor = torch.from_numpy(np.load(config.JOINT_REGRESSOR_H36M)).float()
save_results = result_file is not None
# Disable shuffling if you want to save the results
if save_results:
shuffle=False
# Create dataloader for the dataset
data_loader = DataLoader(dataset, batch_size=batch_size, shuffle=shuffle, num_workers=num_workers)
# Pose metrics
# MPJPE and Reconstruction error for the non-parametric and parametric shapes
# mpjpe = np.zeros(len(dataset))
# recon_err = np.zeros(len(dataset))
quant_mpjpe = {}#np.zeros(len(dataset))
quant_recon_err = {}#np.zeros(len(dataset))
mpjpe = np.zeros(len(dataset))
recon_err = np.zeros(len(dataset))
mpjpe_smpl = np.zeros(len(dataset))
recon_err_smpl = np.zeros(len(dataset))
# Shape metrics
# Mean per-vertex error
shape_err = np.zeros(len(dataset))
shape_err_smpl = np.zeros(len(dataset))
# Mask and part metrics
# Accuracy
accuracy = 0.
parts_accuracy = 0.
# True positive, false positive and false negative
tp = np.zeros((2,1))
fp = np.zeros((2,1))
fn = np.zeros((2,1))
parts_tp = np.zeros((7,1))
parts_fp = np.zeros((7,1))
parts_fn = np.zeros((7,1))
# Pixel count accumulators
pixel_count = 0
parts_pixel_count = 0
# Store SMPL parameters
smpl_pose = np.zeros((len(dataset), 72))
smpl_betas = np.zeros((len(dataset), 10))
smpl_camera = np.zeros((len(dataset), 3))
pred_joints = np.zeros((len(dataset), 17, 3))
eval_pose = False
eval_masks = False
eval_parts = False
# Choose appropriate evaluation for each dataset
if dataset_name == 'h36m-p1' or dataset_name == 'h36m-p2' or dataset_name == 'lspet-test' \
or dataset_name == '3dpw' or dataset_name == 'coco2014-val-3d-amt' or dataset_name == 'ochuman-test' \
or dataset_name == '3dpw-vibe' or dataset_name == '3dpw-crop' or dataset_name == '3dpw-headcrop' or dataset_name == 'mpi-inf-3dhp-test':
eval_pose = True
elif dataset_name == 'lsp':
eval_masks = True
eval_parts = True
annot_path = config.DATASET_FOLDERS['upi-s1h']
joint_mapper_h36m = constants.H36M_TO_J17 if dataset_name == 'mpi-inf-3dhp-test' else constants.H36M_TO_J14
joint_mapper_gt = constants.J24_TO_J17 if dataset_name == 'mpi-inf-3dhp-test' else constants.J24_TO_J14
# Iterate over the entire dataset
for step, batch in enumerate(tqdm(data_loader, desc='Eval', total=len(data_loader))):
# Get ground truth annotations from the batch
imgName = batch['imgname'][0]
seqName = os.path.basename ( os.path.dirname(imgName) )
gt_pose = batch['pose'].to(device)
gt_betas = batch['betas'].to(device)
gt_vertices = smpl_neutral(betas=gt_betas, body_pose=gt_pose[:, 3:], global_orient=gt_pose[:, :3]).vertices
images = batch['img'].to(device)
gender = batch['gender'].to(device)
curr_batch_size = images.shape[0]
# gt_bbox_scale = batch['scale'].cpu().numpy()
# gt_bbox_center = batch['center'].cpu().numpy()
with torch.no_grad():
pred_rotmat, pred_betas, pred_camera = model(images)
pred_output = smpl_neutral(betas=pred_betas, body_pose=pred_rotmat[:,1:], global_orient=pred_rotmat[:,0].unsqueeze(1), pose2rot=False)
pred_vertices = pred_output.vertices
if save_results:
rot_pad = torch.tensor([0,0,1], dtype=torch.float32, device=device).view(1,3,1)
rotmat = torch.cat((pred_rotmat.view(-1, 3, 3), rot_pad.expand(curr_batch_size * 24, -1, -1)), dim=-1)
pred_pose = tgm.rotation_matrix_to_angle_axis(rotmat).contiguous().view(-1, 72)
smpl_pose[step * batch_size:step * batch_size + curr_batch_size, :] = pred_pose.cpu().numpy()
smpl_betas[step * batch_size:step * batch_size + curr_batch_size, :] = pred_betas.cpu().numpy()
smpl_camera[step * batch_size:step * batch_size + curr_batch_size, :] = pred_camera.cpu().numpy()
# 3D pose evaluation
if eval_pose:
# Regressor broadcasting
J_regressor_batch = J_regressor[None, :].expand(pred_vertices.shape[0], -1, -1).to(device)
# Get 14 ground truth joints
if 'h36m' in dataset_name or 'mpi-inf' in dataset_name:
gt_keypoints_3d = batch['pose_3d'].cuda()
gt_keypoints_3d = gt_keypoints_3d[:, joint_mapper_gt, :-1]
# For 3DPW get the 14 common joints from the rendered shape
else:
gt_vertices = smpl_male(global_orient=gt_pose[:,:3], body_pose=gt_pose[:,3:], betas=gt_betas).vertices
gt_vertices_female = smpl_female(global_orient=gt_pose[:,:3], body_pose=gt_pose[:,3:], betas=gt_betas).vertices
if seqName=='val2014':
gt_vertices_neutral = smpl_neutral(global_orient=gt_pose[:,:3], body_pose=gt_pose[:,3:], betas=gt_betas).vertices
gt_vertices = gt_vertices_neutral
else:
gt_vertices[gender==1, :, :] = gt_vertices_female[gender==1, :, :]
gt_keypoints_3d = torch.matmul(J_regressor_batch, gt_vertices)
gt_pelvis = gt_keypoints_3d[:, [0],:].clone()
gt_keypoints_3d = gt_keypoints_3d[:, joint_mapper_h36m, :]
gt_keypoints_3d = gt_keypoints_3d - gt_pelvis
# Get 14 predicted joints from the mesh
pred_keypoints_3d = torch.matmul(J_regressor_batch, pred_vertices)
if save_results:
pred_joints[step * batch_size:step * batch_size + curr_batch_size, :, :] = pred_keypoints_3d.cpu().numpy()
pred_pelvis = pred_keypoints_3d[:, [0],:].clone()
pred_keypoints_3d = pred_keypoints_3d[:, joint_mapper_h36m, :]
pred_keypoints_3d = pred_keypoints_3d - pred_pelvis
# Absolute error (MPJPE)
error = torch.sqrt(((pred_keypoints_3d - gt_keypoints_3d) ** 2).sum(dim=-1)).mean(dim=-1).cpu().numpy()
error_upper = torch.sqrt(((pred_keypoints_3d - gt_keypoints_3d) ** 2).sum(dim=-1)).mean(dim=-1).cpu().numpy()
# mpjpe[step * batch_size:step * batch_size + curr_batch_size] = error
# Reconstuction_error
r_error = reconstruction_error(pred_keypoints_3d.cpu().numpy(), gt_keypoints_3d.cpu().numpy(), reduction=None)
r_error_upper = reconstruction_error(pred_keypoints_3d.cpu().numpy(), gt_keypoints_3d.cpu().numpy(), reduction=None)
# recon_err[step * batch_size:step * batch_size + curr_batch_size] = r_error
#Visualize GT vs prediction
if False:
from renderer import viewer2D
from renderer import glViewer
import humanModelViewer
gt_cam_param = batch['cam_param'].cpu().numpy()
pred_cam_param = pred_camera.detach().cpu().numpy()
batchNum = gt_pose.shape[0]
for i in range(batchNum):
curImgVis = deNormalizeBatchImg(images[i].cpu())
viewer2D.ImShow(curImgVis, name='rawIm', scale=1.0)
#move mesh to bbox space
vert_gt = gt_vertices[i].cpu().numpy()
vert_gt = convert_smpl_to_bbox(vert_gt, gt_cam_param[i][0], gt_cam_param[i][1:])
vert_pred = pred_vertices[i].cpu().numpy()
vert_pred = convert_smpl_to_bbox(vert_pred, pred_cam_param[i][0], pred_cam_param[i][1:])
smpl_face = humanModelViewer.GetSMPLFace()
# meshes_gt = {'ver': gt_vertices[i].cpu().numpy()*100, 'f': smpl_face, 'color': (255,0,0)}
# meshes_pred = {'ver': pred_vertices[i].cpu().numpy()*100, 'f': smpl_face, 'color': (0,255,0)}
meshes_gt = {'ver': vert_gt, 'f': smpl_face, 'color': (200,50,50)}
meshes_pred = {'ver': vert_pred, 'f': smpl_face, 'color': (50,200,50)}
glViewer.setMeshData([meshes_gt, meshes_pred], bComputeNormal= True)
glViewer.setBackgroundTexture(curImgVis) #Vis raw video as background
glViewer.setWindowSize(curImgVis.shape[1]*5, curImgVis.shape[0]*5)
glViewer.SetOrthoCamera(True)
glViewer.show(0)
for ii, p in enumerate(batch['imgname'][:len(r_error)]):
seqName = os.path.basename( os.path.dirname(p))
# quant_mpjpe[step * batch_size:step * batch_size + curr_batch_size] = error
if seqName not in quant_mpjpe.keys():
quant_mpjpe[seqName] = []
quant_recon_err[seqName] = []
quant_mpjpe[seqName].append(error[ii])
quant_recon_err[seqName].append(r_error[ii])
#Visualize GT mesh and Pred Sekeleton
if False:
from renderer import viewer2D
from renderer import glViewer
import humanModelViewer
gt_keypoints_3d_vis = gt_keypoints_3d.cpu().numpy()
gt_keypoints_3d_vis = np.reshape(gt_keypoints_3d_vis, (gt_keypoints_3d_vis.shape[0],-1)) #N,14x3
gt_keypoints_3d_vis = np.swapaxes(gt_keypoints_3d_vis, 0,1) *100
pred_keypoints_3d_vis = pred_keypoints_3d.cpu().numpy()
pred_keypoints_3d_vis = np.reshape(pred_keypoints_3d_vis, (pred_keypoints_3d_vis.shape[0],-1)) #N,14x3
pred_keypoints_3d_vis = np.swapaxes(pred_keypoints_3d_vis, 0,1) *100
# output_sample = output_sample[ : , np.newaxis]*0.1
# gt_sample = gt_sample[: , np.newaxis]*0.1
# (skelNum, dim, frames)
for f in range(gt_keypoints_3d_vis.shape[1]):
glViewer.setSkeleton( [gt_keypoints_3d_vis[:,[f]], pred_keypoints_3d_vis[:,[f]]] ,jointType='smplcoco')#(skelNum, dim, frames)
glViewer.show(0)
# Reconstuction_error
# quant_recon_err[step * batch_size:step * batch_size + curr_batch_size] = r_error
list_mpjpe = np.hstack([ quant_mpjpe[k] for k in quant_mpjpe])
list_reconError = np.hstack([ quant_recon_err[k] for k in quant_recon_err])
if bVerbose:
print(">>> {} : MPJPE {:.02f} mm, error: {:.02f} mm | Total MPJPE {:.02f} mm, error {:.02f} mm".format(seqName, np.mean(error)*1000, np.mean(r_error)*1000, np.hstack(list_mpjpe).mean()*1000, np.hstack(list_reconError).mean()*1000) )
# print("MPJPE {}, error: {}".format(np.mean(error)*100, np.mean(r_error)*100))
# If mask or part evaluation, render the mask and part images
# if eval_masks or eval_parts:
# mask, parts = renderer(pred_vertices, pred_camera)
# Mask evaluation (for LSP)
if eval_masks:
center = batch['center'].cpu().numpy()
scale = batch['scale'].cpu().numpy()
# Dimensions of original image
orig_shape = batch['orig_shape'].cpu().numpy()
for i in range(curr_batch_size):
# After rendering, convert imate back to original resolution
pred_mask = uncrop(mask[i].cpu().numpy(), center[i], scale[i], orig_shape[i]) > 0
# Load gt mask
gt_mask = cv2.imread(os.path.join(annot_path, batch['maskname'][i]), 0) > 0
# Evaluation consistent with the original UP-3D code
accuracy += (gt_mask == pred_mask).sum()
pixel_count += np.prod(np.array(gt_mask.shape))
for c in range(2):
cgt = gt_mask == c
cpred = pred_mask == c
tp[c] += (cgt & cpred).sum()
fp[c] += (~cgt & cpred).sum()
fn[c] += (cgt & ~cpred).sum()
f1 = 2 * tp / (2 * tp + fp + fn)
# Part evaluation (for LSP)
if eval_parts:
center = batch['center'].cpu().numpy()
scale = batch['scale'].cpu().numpy()
orig_shape = batch['orig_shape'].cpu().numpy()
for i in range(curr_batch_size):
pred_parts = uncrop(parts[i].cpu().numpy().astype(np.uint8), center[i], scale[i], orig_shape[i])
# Load gt part segmentation
gt_parts = cv2.imread(os.path.join(annot_path, batch['partname'][i]), 0)
# Evaluation consistent with the original UP-3D code
# 6 parts + background
for c in range(7):
cgt = gt_parts == c
cpred = pred_parts == c
cpred[gt_parts == 255] = 0
parts_tp[c] += (cgt & cpred).sum()
parts_fp[c] += (~cgt & cpred).sum()
parts_fn[c] += (cgt & ~cpred).sum()
gt_parts[gt_parts == 255] = 0
pred_parts[pred_parts == 255] = 0
parts_f1 = 2 * parts_tp / (2 * parts_tp + parts_fp + parts_fn)
parts_accuracy += (gt_parts == pred_parts).sum()
parts_pixel_count += np.prod(np.array(gt_parts.shape))
# Print intermediate results during evaluation
if bVerbose:
if step % log_freq == log_freq - 1:
if eval_pose:
print('MPJPE: ' + str(1000 * mpjpe[:step * batch_size].mean()))
print('Reconstruction Error: ' + str(1000 * recon_err[:step * batch_size].mean()))
print()
if eval_masks:
print('Accuracy: ', accuracy / pixel_count)
print('F1: ', f1.mean())
print()
if eval_parts:
print('Parts Accuracy: ', parts_accuracy / parts_pixel_count)
print('Parts F1 (BG): ', parts_f1[[0,1,2,3,4,5,6]].mean())
print()
# if step==3: #Debug
# break
# Save reconstructions to a file for further processing
if save_results:
np.savez(result_file, pred_joints=pred_joints, pose=smpl_pose, betas=smpl_betas, camera=smpl_camera)
# Print final results during evaluation
if bVerbose:
print('*** Final Results ***')
print()
evalLog ={}
if eval_pose:
# if bVerbose:
# print('MPJPE: ' + str(1000 * mpjpe.mean()))
# print('Reconstruction Error: ' + str(1000 * recon_err.mean()))
# print()
list_mpjpe = np.hstack([ quant_mpjpe[k] for k in quant_mpjpe])
list_reconError = np.hstack([ quant_recon_err[k] for k in quant_recon_err])
output_str ='SeqNames; '
for seq in quant_mpjpe:
output_str += seq + ';'
output_str +='\n MPJPE; '
quant_mpjpe_avg_mm = np.hstack(list_mpjpe).mean()*1000
output_str += "Avg {:.02f} mm; ".format( quant_mpjpe_avg_mm)
for seq in quant_mpjpe:
output_str += '{:.02f}; '.format(1000 * np.hstack(quant_mpjpe[seq]).mean())
output_str +='\n Recon Error; '
quant_recon_error_avg_mm = np.hstack(list_reconError).mean()*1000
output_str +="Avg {:.02f}mm; ".format( quant_recon_error_avg_mm )
for seq in quant_recon_err:
output_str += '{:.02f}; '.format(1000 * np.hstack(quant_recon_err[seq]).mean())
if bVerbose:
print(output_str)
else:
print(">>> Test on 3DPW: MPJPE: {} | quant_recon_error_avg_mm: {}".format(quant_mpjpe_avg_mm, quant_recon_error_avg_mm) )
#Save output to dict
# evalLog['checkpoint']= args.checkpoint
evalLog['testdb'] = dataset_name
evalLog['datasize'] = len(data_loader.dataset)
for seq in quant_mpjpe:
quant_mpjpe[seq] = 1000 * np.hstack(quant_mpjpe[seq]).mean()
for seq in quant_recon_err:
quant_recon_err[seq] = 1000 * np.hstack(quant_recon_err[seq]).mean()
evalLog['quant_mpjpe'] = quant_mpjpe #MPJPE
evalLog['quant_recon_err']= quant_recon_err #PA-MPJPE
evalLog['quant_output_logstr']= output_str #PA-MPJPE
evalLog['quant_mpjpe_avg_mm'] = quant_mpjpe_avg_mm #MPJPE
evalLog['quant_recon_error_avg_mm']= quant_recon_error_avg_mm #PA-MPJPE
# return quant_mpjpe_avg_mm, quant_recon_error_avg_mm, evalLog
return evalLog
if bVerbose:
if eval_masks:
print('Accuracy: ', accuracy / pixel_count)
print('F1: ', f1.mean())
print()
if eval_parts:
print('Parts Accuracy: ', parts_accuracy / parts_pixel_count)
print('Parts F1 (BG): ', parts_f1[[0,1,2,3,4,5,6]].mean())
print()
return -1 #Should return something
def visEFT_singleSubject(renderer):
bStopForEachSample = args.waitforkeys #if True, it will wait for any key pressed to move to the next sample
inputData = args.fit_data
imgDir = args.img_dir
#Load SMPL model
smplModelPath = args.smpl_dir + '/basicModel_neutral_lbs_10_207_0_v1.0.0.pkl'
smpl = SMPL(smplModelPath, batch_size=1, create_transl=False)
print("Loading coco annotation from:{}".format(args.cocoAnnotFile))
assert os.path.exists(args.cocoAnnotFile)
cocoAnnotDic = loadCOCOAnnot(args.cocoAnnotFile)
#Load EFT fitting data
print(f"Loading EFT data from {inputData}")
if os.path.exists(inputData):
with open(inputData, 'r') as f:
eft_data = json.load(f)
print("EFT data: ver {}".format(eft_data['ver']))
eft_data_all = eft_data['data']
else:
print(f"ERROR:: Cannot find EFT data: {inputData}")
assert False
#Visualize each EFT Fitting output
for idx, eft_data in enumerate(eft_data_all):
#Get raw image path
imgFullPath = eft_data['imageName']
imgName = os.path.basename(imgFullPath)
imgFullPath = os.path.join(imgDir, imgName)
if os.path.exists(imgFullPath) == False:
print(f"Img path is not valid: {imgFullPath}")
assert False
rawImg = cv2.imread(imgFullPath)
print(f'Input image: {imgFullPath}')
#EFT data
bbox_scale = eft_data['bbox_scale']
bbox_center = eft_data['bbox_center']
pred_camera = np.array(eft_data['parm_cam'])
pred_betas = np.reshape(
np.array(eft_data['parm_shape'], dtype=np.float32),
(1, 10)) #(10,)
pred_betas = torch.from_numpy(pred_betas)
pred_pose_rotmat = np.reshape(
np.array(eft_data['parm_pose'], dtype=np.float32),
(1, 24, 3, 3)) #(24,3,3)
pred_pose_rotmat = torch.from_numpy(pred_pose_rotmat)
keypoint_2d_validity = eft_data['joint_validity_openpose18']
#COCO only. Annotation index
print("COCO annotId: {}".format(eft_data['annotId']))
annot = cocoAnnotDic[eft_data['annotId']]
print(annot['bbox'])
########################
#Visualize COCO annotation
annot_keypoint = np.reshape(
np.array(annot['keypoints'], dtype=np.float32), (-1, 3)) #17,3
rawImg = viewer2D.Vis_Skeleton_2D_coco(annot_keypoint[:, :2],
annot_keypoint[:, 2],
image=rawImg)
rawImg = viewer2D.Vis_Bbox(rawImg, annot['bbox'], color=(0, 255, 0))
#Get SMPL mesh and joints from SMPL parameters
smpl_output = smpl(betas=pred_betas,
body_pose=pred_pose_rotmat[:, 1:],
global_orient=pred_pose_rotmat[:, [0]],
pose2rot=False)
smpl_vertices = smpl_output.vertices.detach().cpu().numpy()[0]
smpl_joints_3d = smpl_output.joints.detach().cpu().numpy()[0]
#Crop image using cropping information
croppedImg, boxScale_o2n, bboxTopLeft = crop_bboxInfo(
rawImg, bbox_center, bbox_scale, (BBOX_IMG_RES, BBOX_IMG_RES))
########################
# Visualization of EFT
########################
# Visualize 2D image
if True:
viewer2D.ImShow(rawImg, name='rawImg',
waitTime=1) #You should press any key
viewer2D.ImShow(croppedImg, name='croppedImg', waitTime=1)
#Convert bbox_center, bbox_scale --> bbox_xyxy
bbox_xyxy = conv_bboxinfo_bboxXYXY(bbox_scale, bbox_center)
img_bbox = viewer2D.Vis_Bbox_minmaxPt(rawImg.copy(), bbox_xyxy[:2],
bbox_xyxy[2:])
viewer2D.ImShow(img_bbox, name='img_bbox', waitTime=1)
# Visualization Mesh
if True:
camParam_scale = pred_camera[0]
camParam_trans = pred_camera[1:]
pred_vert_vis = smpl_vertices
smpl_joints_3d_vis = smpl_joints_3d
if True: #args.onbbox:
pred_vert_vis = convert_smpl_to_bbox(pred_vert_vis,
camParam_scale,
camParam_trans)
smpl_joints_3d_vis = convert_smpl_to_bbox(
smpl_joints_3d_vis, camParam_scale, camParam_trans)
renderer.setBackgroundTexture(croppedImg)
renderer.setViewportSize(croppedImg.shape[1],
croppedImg.shape[0])
pred_meshes = {'ver': pred_vert_vis, 'f': smpl.faces}
v = pred_meshes['ver']
f = pred_meshes['f']
#Visualize in the original image space
renderer.set_mesh(v, f)
renderer.showBackground(True)
renderer.setWorldCenterBySceneCenter()
renderer.setCameraViewMode("cam")
renderer.setViewportSize(croppedImg.shape[1], croppedImg.shape[0])
renderer.display()
renderImg = renderer.get_screen_color_ibgr()
viewer2D.ImShow(renderImg, waitTime=1)
# Visualization Mesh on side view
if True:
renderer.showBackground(False)
renderer.setWorldCenterBySceneCenter()
renderer.setCameraViewMode("side")
renderer.setViewportSize(croppedImg.shape[1], croppedImg.shape[0])
renderer.display()
sideImg = renderer.get_screen_color_ibgr() #Overwite on rawImg
viewer2D.ImShow(sideImg, waitTime=1)
sideImg = cv2.resize(sideImg,
(renderImg.shape[1], renderImg.shape[0]))
#Visualize camera view and side view
saveImg = np.concatenate((renderImg, sideImg), axis=1)
if bStopForEachSample:
viewer2D.ImShow(
saveImg, waitTime=0
) #waitTime=0 means that it will wait for any key pressed
else:
viewer2D.ImShow(saveImg, waitTime=1)
#Save the rendered image to files
if False:
if os.path.exists(args.render_dir) == False:
os.mkdir(args.render_dir)
render_output_path = args.render_dir + '/render_{:08d}.jpg'.format(
idx)
print(f"Save to {render_output_path}")
cv2.imwrite(render_output_path, saveImg)
def visEFT_multiSubjects(renderer):
bStopForEachSample = args.waitforkeys #if True, it will wait for any key pressed to move to the next sample
bShowTurnTable = args.turntable
# inputDir = args.fit_dir
inputData = args.fit_data
imgDir = args.img_dir
smplModelPath = args.smpl_dir + '/basicModel_neutral_lbs_10_207_0_v1.0.0.pkl'
smpl = SMPL(smplModelPath, batch_size=1, create_transl=False)
if os.path.exists(inputData):
with open(inputData,'r') as f:
eft_data = json.load(f)
print("EFT data: ver {}".format(eft_data['ver']))
eft_data_all = eft_data['data']
else:
print(f"ERROR:: Cannot find EFT data: {inputData}")
assert False
#Aggregate all efl per image
eft_perimage ={}
for idx, eft_data in enumerate(eft_data_all):
#Load
imageName = eft_data['imageName']
if imageName not in eft_perimage.keys():
eft_perimage[imageName] =[]
eft_perimage[imageName].append(eft_data)
for imgName in eft_perimage:
eft_data_perimage = eft_perimage[imgName]
renderer.clear_mesh()
for idx,eft_data in enumerate(eft_data_perimage):
#Get raw image path
imgFullPath = eft_data['imageName']
imgName = os.path.basename(imgFullPath)
imgFullPath =os.path.join(imgDir, imgName)
if os.path.exists(imgFullPath) ==False:
print(f"Img path is not valid: {imgFullPath}")
assert False
rawImg = cv2.imread(imgFullPath)
print(f'Input image: {imgFullPath}')
bbox_scale = eft_data['bbox_scale']
bbox_center = eft_data['bbox_center']
pred_camera = np.array(eft_data['parm_cam'])
pred_betas = np.reshape(np.array( eft_data['parm_shape'], dtype=np.float32), (1,10) ) #(10,)
pred_betas = torch.from_numpy(pred_betas)
pred_pose_rotmat = np.reshape( np.array( eft_data['parm_pose'], dtype=np.float32), (1,24,3,3) ) #(24,3,3)
pred_pose_rotmat = torch.from_numpy(pred_pose_rotmat)
# gt_keypoint_2d = np.reshape( np.array(eft_data['gt_keypoint_2d']), (-1,3)) #(49,3)
keypoint_2d_validity = eft_data['joint_validity_openpose18']
#COCO only. Annotation index
print("COCO annotId: {}".format(eft_data['annotId']))
#Obtain skeleton and smpl data
smpl_output = smpl(betas=pred_betas, body_pose=pred_pose_rotmat[:,1:], global_orient=pred_pose_rotmat[:,0].unsqueeze(1), pose2rot=False )
smpl_vertices = smpl_output.vertices.detach().cpu().numpy()
smpl_joints_3d = smpl_output.joints.detach().cpu().numpy()
#Crop image
croppedImg, boxScale_o2n, bboxTopLeft = crop_bboxInfo(rawImg.copy(), bbox_center, bbox_scale, (BBOX_IMG_RES, BBOX_IMG_RES) )
########################
# Visualize
# Visualize 2D image
if False:
viewer2D.ImShow(rawImg, name='rawImg', waitTime=1) #You should press any key
viewer2D.ImShow(croppedImg, name='croppedImg', waitTime=1)
# Visualization Mesh on raw images
if True:
camParam_scale = pred_camera[0]
camParam_trans = pred_camera[1:]
pred_vert_vis = smpl_vertices[0]
smpl_joints_3d_vis = smpl_joints_3d[0]
if False:#args.onbbox: #Always in the original image
pred_vert_vis = convert_smpl_to_bbox(pred_vert_vis, camParam_scale, camParam_trans)
smpl_joints_3d_vis = convert_smpl_to_bbox(smpl_joints_3d_vis, camParam_scale, camParam_trans)
renderer.setBackgroundTexture(croppedImg)
renderer.setViewportSize(croppedImg.shape[1], croppedImg.shape[0])
else:
#Covert SMPL to BBox first
pred_vert_vis = convert_smpl_to_bbox(pred_vert_vis, camParam_scale, camParam_trans)
smpl_joints_3d_vis = convert_smpl_to_bbox(smpl_joints_3d_vis, camParam_scale, camParam_trans)
#From cropped space to original
pred_vert_vis = convert_bbox_to_oriIm(pred_vert_vis, boxScale_o2n, bboxTopLeft, rawImg.shape[1], rawImg.shape[0])
smpl_joints_3d_vis = convert_bbox_to_oriIm(smpl_joints_3d_vis, boxScale_o2n, bboxTopLeft, rawImg.shape[1], rawImg.shape[0])
renderer.setBackgroundTexture(rawImg)
renderer.setViewportSize(rawImg.shape[1], rawImg.shape[0])
pred_meshes = {'ver': pred_vert_vis, 'f': smpl.faces}
v = pred_meshes['ver']
f = pred_meshes['f']
#Visualize in the original image spaceq
# renderer.set_mesh(v,f)
renderer.add_mesh(v,f)
#Render Mesh on the camera view
renderer.showBackground(True)
renderer.setWorldCenterBySceneCenter()
renderer.setCameraViewMode("cam")
renderer.display()
overlaid = renderer.get_screen_color_ibgr() #Overwite on rawImg
# viewer2D.ImShow(overlaid,waitTime=1,name="overlaid")
if bStopForEachSample:
viewer2D.ImShow(overlaid,waitTime=0,name="overlaid") #waitTime=0 means that it will wait for any key pressed
else:
viewer2D.ImShow(overlaid,waitTime=1,name="overlaid")
#Render Mesh on the rotating view
if bShowTurnTable:
renderer.showBackground(False)
renderer.setWorldCenterBySceneCenter()
renderer.setCameraViewMode("free")
for i in range(90):
renderer.setViewAngle(i*4,0)
renderer.display()
sideImg = renderer.get_screen_color_ibgr() #Overwite on rawImg
viewer2D.ImShow(sideImg,waitTime=1,name="turn_table")
if True: #Save the rendered image to files
if os.path.exists(args.render_dir) == False:
os.mkdir(args.render_dir)
render_output_path = args.render_dir + '/render_{}.jpg'.format(imgName)
print(f"Save to {render_output_path}")
cv2.imwrite(render_output_path, rawImg)
def visEFT_singleSubject(renderer):
MAGNIFY_RATIO = 3 #onbbox only. To magnify the rendered image size
bStopForEachSample = args.waitforkeys #if True, it will wait for any key pressed to move to the next sample
bShowTurnTable = args.turntable
inputData = args.fit_data
imgDir = args.img_dir
#Load SMPL model
smplModelPath = args.smpl_dir + '/basicModel_neutral_lbs_10_207_0_v1.0.0.pkl'
smpl = SMPL(smplModelPath, batch_size=1, create_transl=False)
#Load EFT fitting data
print(f"Loading EFT data from {inputData}")
if os.path.exists(inputData):
with open(inputData, 'r') as f:
eft_data = json.load(f)
print("EFT data: ver {}".format(eft_data['ver']))
eft_data_all = eft_data['data']
else:
print(f"ERROR:: Cannot find EFT data: {inputData}")
assert False
#Visualize each EFT Fitting output
for idx, eft_data in enumerate(tqdm(eft_data_all)):
#Get raw image path
imgFullPath = eft_data['imageName']
# imgName = os.path.basename(imgFullPath)
imgName = imgFullPath
imgFullPath = os.path.join(imgDir, imgName)
if os.path.exists(imgFullPath) == False:
print(f"Img path is not valid: {imgFullPath}")
assert False
rawImg = cv2.imread(imgFullPath)
print(f'Input image: {imgFullPath}')
#EFT data
bbox_scale = eft_data['bbox_scale']
bbox_center = eft_data['bbox_center']
pred_camera = np.array(eft_data['parm_cam'])
pred_betas = np.reshape(
np.array(eft_data['parm_shape'], dtype=np.float32),
(1, 10)) #(10,)
pred_betas = torch.from_numpy(pred_betas)
pred_pose_rotmat = np.reshape(
np.array(eft_data['parm_pose'], dtype=np.float32),
(1, 24, 3, 3)) #(24,3,3)
pred_pose_rotmat = torch.from_numpy(pred_pose_rotmat)
keypoint_2d_validity = eft_data['joint_validity_openpose18']
#COCO only. Annotation index
if 'annotId' in eft_data.keys():
print("COCO annotId: {}".format(eft_data['annotId']))
#Get SMPL mesh and joints from SMPL parameters
smpl_output = smpl(betas=pred_betas,
body_pose=pred_pose_rotmat[:, 1:],
global_orient=pred_pose_rotmat[:, [0]],
pose2rot=False)
smpl_vertices = smpl_output.vertices.detach().cpu().numpy()[0]
smpl_joints_3d = smpl_output.joints.detach().cpu().numpy()[0]
#Crop image using cropping information
croppedImg, boxScale_o2n, bboxTopLeft = crop_bboxInfo(
rawImg, bbox_center, bbox_scale, (BBOX_IMG_RES, BBOX_IMG_RES))
if MAGNIFY_RATIO > 1:
croppedImg = cv2.resize(croppedImg,
(croppedImg.shape[1] * MAGNIFY_RATIO,
croppedImg.shape[0] * MAGNIFY_RATIO))
########################
# Visualization
########################
# Visualize 2D image
if True:
viewer2D.ImShow(rawImg, name='rawImg',
waitTime=1) #You should press any key
viewer2D.ImShow(croppedImg, name='croppedImg', waitTime=1)
#Convert bbox_center, bbox_scale --> bbox_xyxy
bbox_xyxy = conv_bboxinfo_bboxXYXY(bbox_scale, bbox_center)
img_bbox = viewer2D.Vis_Bbox_minmaxPt(rawImg.copy(), bbox_xyxy[:2],
bbox_xyxy[2:])
viewer2D.ImShow(img_bbox, name='img_bbox', waitTime=1)
# Visualization Mesh
if True:
camParam_scale = pred_camera[0]
camParam_trans = pred_camera[1:]
pred_vert_vis = smpl_vertices
smpl_joints_3d_vis = smpl_joints_3d
if args.onbbox:
pred_vert_vis = convert_smpl_to_bbox(pred_vert_vis,
camParam_scale,
camParam_trans)
smpl_joints_3d_vis = convert_smpl_to_bbox(
smpl_joints_3d_vis, camParam_scale, camParam_trans)
renderer.setBackgroundTexture(croppedImg)
renderer.setViewportSize(croppedImg.shape[1],
croppedImg.shape[0])
pred_vert_vis *= MAGNIFY_RATIO
else:
#Covert SMPL to BBox first
pred_vert_vis = convert_smpl_to_bbox(pred_vert_vis,
camParam_scale,
camParam_trans)
smpl_joints_3d_vis = convert_smpl_to_bbox(
smpl_joints_3d_vis, camParam_scale, camParam_trans)
#From cropped space to original
pred_vert_vis = convert_bbox_to_oriIm(pred_vert_vis,
boxScale_o2n,
bboxTopLeft,
rawImg.shape[1],
rawImg.shape[0])
smpl_joints_3d_vis = convert_bbox_to_oriIm(
smpl_joints_3d_vis, boxScale_o2n, bboxTopLeft,
rawImg.shape[1], rawImg.shape[0])
renderer.setBackgroundTexture(rawImg)
renderer.setViewportSize(rawImg.shape[1], rawImg.shape[0])
#In orthographic model. XY of 3D is just 2D projection
smpl_joints_2d_vis = conv_3djoint_2djoint(
smpl_joints_3d_vis, rawImg.shape)
# image_2dkeypoint_pred = viewer2D.Vis_Skeleton_2D_smpl45(smpl_joints_2d_vis, image=rawImg.copy(),color=(0,255,255))
image_2dkeypoint_pred = viewer2D.Vis_Skeleton_2D_Openpose18(
smpl_joints_2d_vis, image=rawImg.copy(),
color=(255, 0, 0)) #All 2D joint
image_2dkeypoint_pred = viewer2D.Vis_Skeleton_2D_Openpose18(
smpl_joints_2d_vis,
pt2d_visibility=keypoint_2d_validity,
image=image_2dkeypoint_pred,
color=(0, 255, 255)) #Only valid
viewer2D.ImShow(image_2dkeypoint_pred,
name='keypoint_2d_pred',
waitTime=1)
pred_meshes = {'ver': pred_vert_vis, 'f': smpl.faces}
v = pred_meshes['ver']
f = pred_meshes['f']
#Visualize in the original image space
renderer.set_mesh(v, f)
renderer.showBackground(True)
renderer.setWorldCenterBySceneCenter()
renderer.setCameraViewMode("cam")
#Set image size for rendering
if args.onbbox:
renderer.setViewportSize(croppedImg.shape[1],
croppedImg.shape[0])
else:
renderer.setViewportSize(rawImg.shape[1], rawImg.shape[0])
renderer.display()
renderImg = renderer.get_screen_color_ibgr()
viewer2D.ImShow(renderImg, waitTime=1)
# Visualization Mesh on side view
if True:
renderer.showBackground(False)
renderer.setWorldCenterBySceneCenter()
# renderer.setCameraViewMode("side") #To show the object in side vie
renderer.setCameraViewMode("free")
renderer.setViewAngle(90, 20)
#Set image size for rendering
if args.onbbox:
renderer.setViewportSize(croppedImg.shape[1],
croppedImg.shape[0])
else:
renderer.setViewportSize(rawImg.shape[1], rawImg.shape[0])
renderer.display()
sideImg = renderer.get_screen_color_ibgr() #Overwite on rawImg
viewer2D.ImShow(sideImg, waitTime=1)
sideImg = cv2.resize(sideImg,
(renderImg.shape[1], renderImg.shape[0]))
# renderImg = cv2.resize(renderImg, (sideImg.shape[1], sideImg.shape[0]) )
# Visualization Mesh on side view
if True:
renderer.showBackground(False)
renderer.setWorldCenterBySceneCenter()
# renderer.setCameraViewMode("side") #To show the object in side vie
renderer.setCameraViewMode("free")
renderer.setViewAngle(-60, 50)
#Set image size for rendering
if args.onbbox:
renderer.setViewportSize(croppedImg.shape[1],
croppedImg.shape[0])
else:
renderer.setViewportSize(rawImg.shape[1], rawImg.shape[0])
renderer.display()
sideImg_2 = renderer.get_screen_color_ibgr() #Overwite on rawImg
viewer2D.ImShow(sideImg_2, waitTime=1)
sideImg_2 = cv2.resize(sideImg_2,
(renderImg.shape[1], renderImg.shape[0]))
# renderImg = cv2.resize(renderImg, (sideImg.shape[1], sideImg.shape[0]) )
#Visualize camera view and side view
saveImg = np.concatenate((renderImg, sideImg), axis=1)
# saveImg = np.concatenate( (croppedImg, renderImg,sideImg, sideImg_2), axis =1)
if bStopForEachSample:
viewer2D.ImShow(
saveImg, waitTime=0
) #waitTime=0 means that it will wait for any key pressed
else:
viewer2D.ImShow(saveImg, waitTime=1)
#Render Mesh on the rotating view
if bShowTurnTable:
renderer.showBackground(False)
renderer.setWorldCenterBySceneCenter()
renderer.setCameraViewMode("free")
for i in range(90):
renderer.setViewAngle(i * 4, 0)
renderer.display()
sideImg = renderer.get_screen_color_ibgr() #Overwite on rawImg
viewer2D.ImShow(sideImg, waitTime=1, name="turn_table")
if False: #If you want to save this into files
render_output_path = args.render_dir + '/turntable_{}_{:08d}.jpg'.format(
os.path.basename(imgName), i)
cv2.imwrite(render_output_path, sideImg)
#Save the rendered image to files
if True:
if os.path.exists(args.render_dir) == False:
os.mkdir(args.render_dir)
render_output_path = args.render_dir + '/render_{}_eft{:08d}.jpg'.format(
imgName[:-4], idx)
print(f"Save to {render_output_path}")
cv2.imwrite(render_output_path, saveImg)