def exportOursToSpin(eftDir, out_path):
# scaleFactor = 1.2
# structs we need
imgnames_, scales_, centers_, parts_, openposes_ = [], [], [], [], []
#additional 3D
poses_, shapes_, skel3D_, has_smpl_ = [], [], [], []
pose3DList = os.listdir(eftDir)
# for imgSample in cocoPose3DAll:
sampleNum = len(pose3DList)
# totalSampleNum = [ len(cocoPose3DAll[imgSample]) for imgSample in cocoPose3DAll ]
# totalSampleNum = sum(totalSampleNum)
print("\n\n### SampleNum: {} ###".format(sampleNum))
maxDiff = 0
for fname in tqdm(sorted(pose3DList)):
fname_path = os.path.join(eftDir, fname)
pose3d = pickle.load(open(fname_path, 'rb'))
#load image
imgPathFull = pose3d['imageName'][0]
fileName = os.path.basename(imgPathFull)
fileName_saved = os.path.join(
os.path.basename(os.path.dirname(imgPathFull)),
fileName) #start from train2014
center = pose3d['center'][0]
scale = pose3d['scale'][0]
smpl_shape = pose3d['pred_shape'].ravel()
smpl_pose_mat = torch.from_numpy(
pose3d['pred_pose_rotmat'][0]) #24,3,3
pred_rotmat_hom = torch.cat([
smpl_pose_mat.view(-1, 3, 3),
torch.tensor(
[0, 0, 0],
dtype=torch.float32,
).view(1, 3, 1).expand(24, -1, -1)
],
dim=-1)
smpl_pose = tgm.rotation_matrix_to_angle_axis(
pred_rotmat_hom).contiguous().view(-1, 72)
#verification
if True:
recon_mat = batch_rodrigues(smpl_pose.view(
-1, 3)) #24,3... axis -> rotmat
diff = abs(recon_mat.numpy() -
pose3d['pred_pose_rotmat'][0]) #2.1234155e-07
# print(np.max(diff))
maxDiff = max(maxDiff, np.max(diff))
smpl_pose = smpl_pose.numpy().ravel()
openpose2d = pose3d['keypoint2d'][0][:25] #25,3
spin2d_skel24 = pose3d['keypoint2d'][0][25:] #24,3
#Save data
imgnames_.append(fileName_saved)
centers_.append(center)
scales_.append(scale)
has_smpl_.append(1)
poses_.append(smpl_pose) #(72,)
shapes_.append(smpl_shape) #(10,)
openposes_.append(openpose2d) #blank
# print(openpose2d)/
parts_.append(spin2d_skel24)
#3D joint
S = np.zeros(
[24, 4]) #blank for 3d. TODO: may need to add valid data for this
skel3D_.append(S)
#Debug 2D Visualize
if False:
img = cv2.imread(
os.path.join('/run/media/hjoo/disk/data/coco', imgnames_[-1]))
img = viewer2D.Vis_Skeleton_2D_smplCOCO(
gt_skel, pt2d_visibility=gt_validity[:, 0], image=img)
img = viewer2D.Vis_Bbox_minmaxPt(img, min_pt, max_pt)
viewer2D.ImShow(img, waitTime=0)
#Debug 3D Visualize smpl_coco
if False:
# data3D_coco_vis = np.reshape(data3D_coco, (data3D_coco.shape[0],-1)).transpose() #(Dim, F)
# data3D_coco_vis *=0.1 #mm to cm
# glViewer.setSkeleton( [ data3D_coco_vis] ,jointType='smplcoco')
# glViewer.show()
#Debug 3D Visualize, h36m
data3D_h36m_vis = np.reshape(
data3D_h36m, (data3D_h36m.shape[0], -1)).transpose() #(Dim, F)
data3D_h36m_vis *= 100 #meter to cm
# data3D_smpl24 = np.reshape(data3D_smpl24, (data3D_smpl24.shape[0],-1)).transpose() #(Dim, F)
# data3D_smpl24 *=0.1
glViewer.setSkeleton([data3D_h36m_vis], jointType='smplcoco')
glViewer.show()
# keypoints
# print("Final Img Num: {}, Final Sample Num: {}".format( len(set(imgnames_) , len(imgnames_)) ) )
print("Final Sample Num: {}".format(len(imgnames_)))
print("maxDiff in rot conv.: {}".format(maxDiff))
# store the data struct
if not os.path.isdir(out_path):
os.makedirs(out_path)
out_file = os.path.join(out_path, os.path.basename(eftDir) + '.npz')
print(f"Save to {out_file}")
np.savez(out_file,
imgname=imgnames_,
center=centers_,
scale=scales_,
part=parts_,
openpose=openposes_,
pose=poses_,
shape=shapes_,
has_smpl=has_smpl_,
S=skel3D_)
def exportOursToSpin(cocoPose3DAll, out_path):
scaleFactor = 1.2
# structs we need
imgnames_, scales_, centers_, parts_, openposes_ = [], [], [], [], []
#additional 3D
poses_ , shapes_, skel3D_, has_smpl_ = [], [] ,[], []
# for imgSample in cocoPose3DAll:
imgNum = len(cocoPose3DAll)
totalSampleNum = [ len(cocoPose3DAll[imgSample]) for imgSample in cocoPose3DAll ]
totalSampleNum = sum(totalSampleNum)
print("\n\n### ImageNum: {}, SampleNum: {} ###".format(imgNum, totalSampleNum))
# for imgSample in cocoPose3DAll:
# for key_imgId, imgSample in sorted(cocoPose3DAll.items()):
for key_imgId, imgSample in sorted(cocoPose3DAll.items()):
#load image
imgPathFull = imgSample[0]['imgId']
fileName = os.path.basename(imgPathFull)
fileName_saved = os.path.join(os.path.basename(os.path.dirname(imgPathFull)), fileName) #start from train2014
for sample in imgSample:
validJointNum = np.sum(sample['pose2D_validity'][::2])
if validJointNum<4:
continue
if np.isnan(sample['pose3DParam']['camScale']):
continue
gt_skel = np.reshape(sample['pose2D_gt'],(26,-1)) #(26,2) This is from data
gt_validity = np.reshape(sample['pose2D_validity'],(26,-1)) #(26,2)
# Filtering ########################################################################################################
if True:
requiredJoints= [0,1,2, 3,4,5, 6,7,8, 9,10,11] #In Total26
if np.min(gt_validity[requiredJoints,0])== False:
continue
min_pt = np.min(gt_skel[gt_validity[:,0]], axis=0)
max_pt = np.max(gt_skel[gt_validity[:,0]], axis=0)
# bbox= [ min_pt[0], min_pt[1], max_pt[0], max_pt[1] ]
bbox= [ min_pt[0], min_pt[1], max_pt[0] - min_pt[0], max_pt[1] - min_pt[1]]
center = [bbox[0] + bbox[2]/2, bbox[1] + bbox[3]/2]
scale = scaleFactor*max(bbox[2], bbox[3])/200
#Save data
imgnames_.append(os.path.join('images',fileName_saved))
openposes_.append(np.zeros([25,3])) #blank
centers_.append(center)
scales_.append(scale)
has_smpl_.append(1)
poses_.append(sample['pose3DParam']['pose']) #(72,)
shapes_.append(sample['pose3DParam']['shape']) #(10,)
#2D keypoints (total26 -> SPIN24)
poseidx_spin24 = [0,1,2, 3,4,5, 6,7,8, 9,10,11, 19,20,21,22,23]
poseidx_total26 = [0,1,2, 3,4,5, 6,7,8, 9,10,11, 14, 15, 16, 17, 18 ]
part = np.zeros([24,3])
part[poseidx_spin24,:2] = gt_skel[poseidx_total26] #(52,) totalGT26 type
part[poseidx_spin24,2] = 1*gt_validity[poseidx_total26,0]
parts_.append(part)
#3D joint
S = np.zeros([24,4])
S[poseidx_spin24,:3] = sample['pose3D_pred'][poseidx_total26,:] * 0.001 #Scaling skeleton 3D (currently mm) -> meter
S[poseidx_spin24,3] = 1
skel3D_.append(S)
#Debug 2D Visualize
if False:
img = cv2.imread( os.path.join( '/run/media/hjoo/disk/data/mpii_human_pose_v1/images',imgnames_[-1]) )
img = viewer2D.Vis_Skeleton_2D_smplCOCO(gt_skel, pt2d_visibility = gt_validity[:,0], image =img)
img = viewer2D.Vis_Bbox_minmaxPt(img, min_pt, max_pt)
viewer2D.ImShow(img, waitTime=0)
#Debug 3D Visualize smpl_coco
if False:
# data3D_coco_vis = np.reshape(data3D_coco, (data3D_coco.shape[0],-1)).transpose() #(Dim, F)
# data3D_coco_vis *=0.1 #mm to cm
# glViewer.setSkeleton( [ data3D_coco_vis] ,jointType='smplcoco')
# glViewer.show()
#Debug 3D Visualize, h36m
data3D_h36m_vis = np.reshape(data3D_h36m, (data3D_h36m.shape[0],-1)).transpose() #(Dim, F)
data3D_h36m_vis *=0.001 #meter to cm
# data3D_smpl24 = np.reshape(data3D_smpl24, (data3D_smpl24.shape[0],-1)).transpose() #(Dim, F)
# data3D_smpl24 *=0.1
glViewer.setSkeleton( [ data3D_h36m_vis] ,jointType='smplcoco')
glViewer.show()
# keypoints
# print("Final Img Num: {}, Final Sample Num: {}".format( len(set(imgnames_) , len(imgnames_)) ) )
print("Final Sample Num: {}".format( len(imgnames_)))
# store the data struct
if not os.path.isdir(out_path):
os.makedirs(out_path)
# out_file = os.path.join(out_path, '1031-mpii3D_train_44257_all.npz')
out_file = os.path.join(out_path, '1031-mpii3D_train_44257_validlimbs.npz')
np.savez(out_file, imgname=imgnames_,
center=centers_,
scale=scales_,
part=parts_,
openpose=openposes_,
pose=poses_,
shape=shapes_,
has_smpl=has_smpl_,
S=skel3D_)
def exportOursToSpin(cocoPose3DAll, out_path):
scaleFactor = 1.2
# annotation files
annot_file = os.path.join('/home/hjoo/codes/SPIN/data', 'train.h5')
# read annotations
f = h5py.File(annot_file, 'r')
centers, imgnames, parts, scales = \
f['center'], f['imgname'], f['part'], f['scale']
# structs we need
imgnames_, scales_, centers_, parts_, openposes_ = [], [], [], [], []
#additional 3D
poses_, shapes_, skel3D_, has_smpl_ = [], [], [], []
# for imgSample in cocoPose3DAll:
imgNum = len(cocoPose3DAll)
totalSampleNum = [
len(cocoPose3DAll[imgSample]) for imgSample in cocoPose3DAll
]
totalSampleNum = sum(totalSampleNum)
print("\n\n### ImageNum: {}, SampleNum: {} ###".format(
imgNum, totalSampleNum))
# for imgSample in cocoPose3DAll:
# for key_imgId, imgSample in sorted(cocoPose3DAll.items()):
for key_imgId, imgSample in sorted(cocoPose3DAll.items()):
#load image
fileName_saved = os.path.join('images', imgSample[0]['imgId'])
# fileName = os.path.basename(imgPathFull)
# fileName_saved = os.path.join(os.path.basename(os.path.dirname(imgPathFull)), fileName) #start from train2014
for sample in imgSample:
validJointNum = np.sum(sample['pose2D_validity'][::2])
if validJointNum < 4:
continue
if np.isnan(sample['pose3DParam']['camScale']):
continue
#visualize
# if True: #Save BBox
# inputImg_2dvis = inputImg.copy()
# inputImg_2dvis = viewer2D.Vis_Bbox(inputImg_2dvis,sample['bbr'])
# # viewer2D.ImShow(inputImg_2dvis)
# imgFilePath = '{0}/{1}.jpg'.format(bboxFolder, fileName)
# cv2.imwrite(imgFilePath, inputImg_2dvis)
if 'bbr' in sample.keys():
bbox = sample['bbr']
else:
keypoints = np.reshape(sample['pose2D_gt'], (-1, 2)) #26,2
valid = sample['pose2D_validity'][::2] #(26,)
valid_keypoints = keypoints[valid]
min_pt = np.min(valid_keypoints, axis=0)
max_pt = np.max(valid_keypoints, axis=0)
bbox = [
min_pt[0], min_pt[1], max_pt[0] - min_pt[0],
max_pt[1] - min_pt[1]
]
#Debug Visualization
if False:
imgPathFull = os.path.join(
'/run/media/hjoo/disk/data/mpii_human_pose_v1',
fileName_saved)
inputImg_2dvis = cv2.imread(imgPathFull)
inputImg_2dvis = viewer2D.Vis_Skeleton_2D_smplCOCO(
sample['pose2D_gt'],
pt2d_visibility=sample['pose2D_validity'],
image=inputImg_2dvis)
inputImg_2dvis = viewer2D.Vis_Bbox(inputImg_2dvis, bbox)
viewer2D.ImShow(inputImg_2dvis, waitTime=0)
# continue
center = [bbox[0] + bbox[2] / 2, bbox[1] + bbox[3] / 2]
scale = scaleFactor * max(bbox[2], bbox[3]) / 200
#Save data
imgnames_.append(fileName_saved)
openposes_.append(np.zeros([25, 3])) #blank
centers_.append(center)
scales_.append(scale)
has_smpl_.append(1)
poses_.append(sample['pose3DParam']['pose']) #(72,)
shapes_.append(sample['pose3DParam']['shape']) #(10,)
#2D keypoints (total26 -> SPIN24)
poseidx_spin24 = [
0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 19, 20, 21, 22, 23, 13
] #13 head top
poseidx_total26 = [
0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 14, 15, 16, 17, 18, 19
] #19 head top
part = np.zeros([24, 3])
gt_skel = np.reshape(sample['pose2D_gt'], (26, -1))
gt_validity = np.reshape(sample['pose2D_validity'], (26, -1))
part[poseidx_spin24, :2] = gt_skel[
poseidx_total26] #(52,) totalGT26 type
part[poseidx_spin24, 2] = 1 * gt_validity[poseidx_total26, 0]
parts_.append(part)
#3D joint
S = np.zeros([24, 4])
S[poseidx_spin24, :3] = sample['pose3D_pred'][poseidx_total26, :]
S[poseidx_spin24, 3] = 1
skel3D_.append(S)
# keypoints
# store the data struct
if not os.path.isdir(out_path):
os.makedirs(out_path)
out_file = os.path.join(out_path,
'mpi_train_wShape_1537_ep200_exemplar.npz')
np.savez(out_file,
imgname=imgnames_,
center=centers_,
scale=scales_,
part=parts_,
openpose=openposes_,
pose=poses_,
shape=shapes_,
has_smpl=has_smpl_,
S=skel3D_)
