def meshPlay(folder, every=100, wait=0.05):
files = glob.glob(folder + '/*')
files.sort()
files = files[-1000:]
view = MeshViewer()
for i in range(0, len(files), every):
mesh = Mesh(filename=files[i])
view.dynamic_meshes = [mesh]
time.sleep(wait)
def optimize_pose_shape(th_scan_meshes,
smpl,
iterations,
steps_per_iter,
th_pose_3d=None,
display=None):
"""
Optimize SMPL.
:param display: if not None, pass index of the scan in th_scan_meshes to visualize.
"""
# Optimizer
optimizer = torch.optim.Adam([smpl.trans, smpl.betas, smpl.pose],
0.02,
betas=(0.9, 0.999))
# Get loss_weights
weight_dict = get_loss_weights()
# Display
if display is not None:
assert int(display) < len(th_scan_meshes)
mv = MeshViewer()
for it in range(iterations):
loop = tqdm(range(steps_per_iter))
loop.set_description('Optimizing SMPL')
for i in loop:
optimizer.zero_grad()
# Get losses for a forward pass
loss_dict = forward_step(th_scan_meshes, smpl, th_pose_3d)
# Get total loss for backward pass
tot_loss = backward_step(loss_dict, weight_dict, it)
tot_loss.backward()
optimizer.step()
l_str = 'Iter: {}'.format(i)
for k in loss_dict:
l_str += ', {}: {:0.4f}'.format(
k, weight_dict[k](loss_dict[k], it).mean().item())
loop.set_description(l_str)
if display is not None:
verts, _, _, _ = smpl()
smpl_mesh = Mesh(v=verts[display].cpu().detach().numpy(),
f=smpl.faces.cpu().numpy())
scan_mesh = Mesh(
v=th_scan_meshes[display].vertices.cpu().detach().numpy(),
f=th_scan_meshes[display].faces.cpu().numpy(),
vc=np.array([0, 1, 0]))
mv.set_static_meshes([scan_mesh, smpl_mesh])
print('** Optimised smpl pose and shape **')
def evaluate(coma, test_loader, dataset, template_mesh, device, visualize,
output_dir):
coma.eval()
total_loss = 0
for i, data in enumerate(test_loader):
data = data.to(device)
with torch.no_grad():
out, mu, logvar = coma(data)
loss = loss_function(out, data.y, mu, logvar)
total_loss += data.num_graphs * loss.item()
if visualize and i % 100 == 0:
meshviewer = MeshViewer(shape=(1, 2))
save_out = out.detach().cpu().numpy()
save_out = save_out * dataset.std.numpy() + dataset.mean.numpy()
expected_out = (data.y.detach().cpu().numpy()
) * dataset.std.numpy() + dataset.mean.numpy()
result_mesh = Mesh(v=save_out, f=template_mesh.f)
expected_mesh = Mesh(v=expected_out, f=template_mesh.f)
meshviewer[0][0].set_dynamic_meshes([result_mesh])
meshviewer[0][1].set_dynamic_meshes([expected_mesh])
meshviewer[0][0].save_snapshot(os.path.join(
output_dir, 'file' + str(i) + '.png'),
blocking=True)
result_mesh.write_ply('{}/result_{}.ply'.format(output_dir, i))
expected_mesh.write_ply('{}/expected_{}.ply'.format(output_dir, i))
print('result mesh and expected mesh are saved as .ply')
return total_loss / len(dataset)
def render_video(mesh_dir, video_fn, overwrite=False):
from psbody.mesh import Mesh, MeshViewer
from os.path import join, exists, splitext
from glob import glob
import tempfile
from subprocess import call
from pickle import load
import numpy as np
from tqdm import tqdm
if exists(video_fn):
if overwrite:
print("File {0} exists, removing it and remaking it".format(
video_fn))
call(['rm', '-rf', video_fn])
else:
print("File {0} exists, not re-rendering".format(video_fn))
return
files_seq = sorted(glob(join(mesh_dir, '*.obj')))
if len(files_seq) == 0:
print('No files to render in {}'.format(mesh_dir))
return
# Load the meshes
print("Loading meshes from {}..".format(mesh_dir))
meshes = []
for fn in files_seq:
meshes.append(Mesh(filename=fn))
from shutil import rmtree
from tempfile import mkdtemp
tmp_folder = str(mkdtemp())
if exists(tmp_folder):
rmtree(tmp_folder)
from os import mkdir
mkdir(tmp_folder)
mv = MeshViewer(window_width=1000, window_height=800)
print('Rendering extracted meshes (tmp file, auto-removed later)..')
for k, mesh in enumerate(tqdm(meshes)):
mv.set_dynamic_meshes([mesh])
mv.save_snapshot(join(tmp_folder, '{:0>6d}.png'.format(k)),
blocking=True)
cmd = [
'ffmpeg', '-i', '{0}/%06d.png'.format(tmp_folder), '-vcodec', 'h264',
'-pix_fmt', 'yuv420p', '-r', '15', '-an', '-b:v', '5000k', video_fn
]
call(cmd)
rmtree(tmp_folder)
def final_fit(
opt,
part_mesh,
v,
v_offset,
dist_o,
dist_i,
smpl_h_ref,
rn_m,
debug_rn,
dif_mask,
v_ids_template,
faces_template,
v_ids_side,
faces_side,
max_y,
proj_cam,
ref_joint_list_coup,
):
if opt.disp_mesh_side:
mv = MeshViewer()
else:
mv = None
if opt.disp_mesh_whl:
mv2 = MeshViewer()
else:
mv2 = None
import scipy.sparse as sp
sparse_solver = lambda A, x: sp.linalg.cg(A, x, maxiter=500)[0]
tgt_pose = get_pose_prior(init_pose_path=opt.init_pose_path,
gar_type=opt.gar_type)
E = {
'mask':
gaussian_pyramid(rn_m * dist_o * opt.ref_wt_dist_o +
(1 - rn_m) * dist_i,
n_levels=4,
normalization='size') * opt.ref_wt_mask
}
x0 = [v_offset]
if opt.ref_wt_coup:
# x0 = [smpl_h_ref.trans, smpl_h_ref.betas, v_offset]
E['coupling'] = (v + v_offset -
smpl_h_ref[v_ids_template]) * opt.ref_wt_coup
if opt.ref_wt_shp:
E['beta_prior'] = ch.linalg.norm(smpl_h_ref.betas) * opt.ref_wt_shp
if opt.ref_wt_pose:
E['pose'] = (smpl_h_ref.pose - tgt_pose) * opt.ref_wt_pose
if ref_joint_list_coup != None:
range_joint = []
for item in ref_joint_list_coup:
range_joint.append(3 * int(item))
range_joint.append(3 * int(item) + 1)
range_joint.append(3 * int(item) + 2)
x0 = x0 + [smpl_h_ref.pose[range_joint]]
if opt.ref_use_betas:
x0 = x0 + [smpl_h_ref.betas]
if opt.ref_wt_proj:
error_bd = get_rings_error(proj_cam, max_y)
E['proj_bd'] = error_bd * opt.ref_wt_proj
if opt.ref_wt_bd:
gar_rings = compute_boundaries(v + v_offset, faces_template)
error = smooth_rings(gar_rings, v + v_offset)
E['boundary'] = error * opt.ref_wt_bd
if opt.ref_wt_lap:
lap_op = np.asarray(laplacian(part_mesh).todense())
lap_err = ch.dot(lap_op, v + v_offset)
E['laplacian'] = lap_err * opt.ref_wt_lap
ch.minimize(E,
x0,
method='dogleg',
options={
'e_3': .000001,
'sparse_solver': sparse_solver
},
callback=get_callback_ref(rend=debug_rn,
mask=dif_mask,
vertices=v + v_offset,
display=opt.display,
v_ids_sides=v_ids_side,
faces_template=faces_template,
faces_side=faces_side,
disp_mesh_side=opt.disp_mesh_side,
disp_mesh_whl=opt.disp_mesh_whl,
save_dir=opt.save_opt_images,
mv=mv,
mv2=mv2,
show=opt.show))
final_mask = rn_m.r
mask = dif_mask.copy()
mask[dif_mask == 0.5] = 1
final_iou = compute_iou(mask, final_mask)
return v + v_offset, final_iou
def init_fit(opt, dist_o, dist_i, dif_mask, rn_m, smpl_h, v_ids_template,
faces_template, debug_rn, v_ids_side, faces_side, joints_list):
range_joint = []
for item in joints_list:
range_joint.append(3 * int(item))
range_joint.append(3 * int(item) + 1)
range_joint.append(3 * int(item) + 2)
tgt_pose = get_pose_prior(init_pose_path=opt.init_pose_path,
gar_type=opt.gar_type)
# ============================================
# FIRST STAGE
# ============================================
from psbody.mesh import MeshViewer
if opt.disp_mesh_side:
mv = MeshViewer()
else:
mv = None
if opt.disp_mesh_whl:
mv2 = MeshViewer()
else:
mv2 = None
if opt.init_first_stage == "Trans":
x0 = [smpl_h.trans]
elif opt.init_first_stage == 'Pose':
x0 = [smpl_h.trans, smpl_h.pose[range_joint]]
# x0 = [smpl_h.trans]
elif opt.init_first_stage == 'Shape':
x0 = [smpl_h.trans, smpl_h.betas]
elif opt.init_first_stage == 'Both':
x0 = [smpl_h.trans, smpl_h.betas, smpl_h.pose[range_joint]]
E = {
'mask':
gaussian_pyramid(rn_m * dist_o * opt.init_fst_wt_dist_o +
(1 - rn_m) * dist_i,
n_levels=4,
normalization='size') * opt.init_fst_wt_mask
}
if opt.init_fst_wt_betas:
E['beta_prior'] = ch.linalg.norm(smpl_h.betas) * opt.init_fst_wt_betas
if opt.init_fst_wt_pose:
E['pose'] = (smpl_h.pose - tgt_pose) * opt.init_fst_wt_pose
ch.minimize(E,
x0,
method='dogleg',
options={
'e_3': .0001,
'disp': True
},
callback=get_callback(rend=debug_rn,
mask=dif_mask,
smpl=smpl_h,
v_ids_template=v_ids_template,
v_ids_sides=v_ids_side,
faces_template=faces_template,
faces_side=faces_side,
display=opt.display,
disp_mesh_side=opt.disp_mesh_side,
disp_mesh_whl=opt.disp_mesh_whl,
mv=mv,
mv2=mv2,
save_dir=opt.save_opt_images,
show=opt.show))
# ===============================================
# SECOND STAGE
# ===============================================
if opt.init_second_stage != "None":
if opt.init_second_stage == 'Pose':
x0 = [smpl_h.trans, smpl_h.pose[range_joint]]
elif opt.init_second_stage == 'Shape':
x0 = [smpl_h.trans, smpl_h.betas]
elif opt.init_second_stage == 'Both':
x0 = [smpl_h.trans, smpl_h.betas, smpl_h.pose[range_joint]]
E = {
'mask':
gaussian_pyramid(rn_m * dist_o * opt.init_sec_wt_dist_o +
(1 - rn_m) * dist_i,
n_levels=4,
normalization='size') * opt.init_sec_wt_mask
}
if opt.init_sec_wt_betas:
E['beta_prior'] = ch.linalg.norm(
smpl_h.betas) * opt.init_sec_wt_betas
if opt.init_sec_wt_pose:
E['pose'] = (smpl_h.pose - tgt_pose) * opt.init_sec_wt_pose
ch.minimize(E,
x0,
method='dogleg',
options={'e_3': .0001},
callback=get_callback(rend=debug_rn,
mask=dif_mask,
smpl=smpl_h,
v_ids_template=v_ids_template,
v_ids_sides=v_ids_side,
faces_template=faces_template,
faces_side=faces_side,
display=opt.display,
disp_mesh_side=opt.disp_mesh_side,
disp_mesh_whl=opt.disp_mesh_whl,
mv=mv,
mv2=mv2,
save_dir=opt.save_opt_images,
show=opt.show))
temp_params = {
'pose': smpl_h.pose.r,
'betas': smpl_h.betas.r,
'trans': smpl_h.trans.r,
'v_personal': smpl_h.v_personal.r
}
part_mesh = Mesh(smpl_h.r[v_ids_template], faces_template)
return part_mesh, temp_params
def optimize_pose_only(th_scan_meshes,
smplx,
iterations,
steps_per_iter,
scan_part_labels,
smplx_part_labels,
search_tree=None,
pen_distance=None,
tri_filtering_module=None,
display=None):
"""
Initially we want to only optimize the global rotation of SMPLX. Next we optimize full pose.
We optimize pose based on the 3D keypoints in th_pose_3d.
:param th_pose_3d: array containing the 3D keypoints.
"""
batch_sz = 1 # smplx.pose.shape[0]
split_smplx = th_batch_SMPLX_split_params(
batch_sz,
top_betas=smplx.betas.data[:, :2],
other_betas=smplx.betas.data[:, 2:],
global_pose=smplx.global_pose.data,
body_pose=smplx.body_pose.data,
left_hand_pose=smplx.left_hand_pose.data,
right_hand_pose=smplx.right_hand_pose.data,
expression=smplx.expression.data,
jaw_pose=smplx.jaw_pose.data,
leye_pose=smplx.leye_pose.data,
reye_pose=smplx.reye_pose.data,
faces=smplx.faces,
gender=smplx.gender).to(DEVICE)
# split_smplx.expression.requires_grad = False
# split_smplx.jaw_pose.requires_grad = False
optimizer = torch.optim.Adam(
[split_smplx.trans, split_smplx.top_betas, split_smplx.global_pose],
0.02,
betas=(0.9, 0.999))
# Get loss_weights
weight_dict = get_loss_weights()
if display is not None:
assert int(display) < len(th_scan_meshes)
# mvs = MeshViewers((1,1))
mv = MeshViewer(keepalive=True)
iter_for_global = 1
for it in range(iter_for_global + iterations):
loop = tqdm(range(steps_per_iter))
if it < iter_for_global:
# Optimize global orientation
print('Optimizing SMPLX global orientation')
loop.set_description('Optimizing SMPLX global orientation')
elif it == iter_for_global:
# Now optimize full SMPLX pose
print('Optimizing SMPLX pose only')
loop.set_description('Optimizing SMPLX pose only')
optimizer = torch.optim.Adam([
split_smplx.trans, split_smplx.top_betas,
split_smplx.global_pose, split_smplx.body_pose,
split_smplx.left_hand_pose, split_smplx.right_hand_pose
],
0.02,
betas=(0.9, 0.999))
else:
loop.set_description('Optimizing SMPLX pose only')
for i in loop:
optimizer.zero_grad()
# Get losses for a forward pass
loss_dict = forward_step(th_scan_meshes, split_smplx,
scan_part_labels, smplx_part_labels,
search_tree, pen_distance,
tri_filtering_module)
# Get total loss for backward pass
tot_loss = backward_step(loss_dict, weight_dict, it)
tot_loss.backward()
optimizer.step()
l_str = 'Iter: {}'.format(i)
for k in loss_dict:
l_str += ', {}: {:0.4f}'.format(
k, weight_dict[k](loss_dict[k], it).mean().item())
loop.set_description(l_str)
if display is not None:
# verts, _, _, _ = split_smplx()
verts = split_smplx()
smplx_mesh = Mesh(v=verts[display].cpu().detach().numpy(),
f=smplx.faces.cpu().numpy())
scan_mesh = Mesh(
v=th_scan_meshes[display].vertices.cpu().detach().numpy(),
f=th_scan_meshes[display].faces.cpu().numpy(),
vc=np.array([0, 1, 0]))
scan_mesh.set_vertex_colors_from_weights(
scan_part_labels[display].cpu().detach().numpy())
mv.set_dynamic_meshes([smplx_mesh, scan_mesh])
# Put back pose, shape and trans into original smplx
smplx.global_pose.data = split_smplx.global_pose.data
smplx.body_pose.data = split_smplx.body_pose.data
smplx.left_hand_pose.data = split_smplx.left_hand_pose.data
smplx.right_hand_pose.data = split_smplx.right_hand_pose.data
# smplx.jaw_pose.data = split_smplx.jaw_pose.data
smplx.leye_pose.data = split_smplx.leye_pose.data
smplx.reye_pose.data = split_smplx.reye_pose.data
smplx.betas.data = split_smplx.betas.data
smplx.trans.data = split_smplx.trans.data
print('** Optimised smplx pose **')
def optimize_pose_shape(th_scan_meshes,
smplx,
iterations,
steps_per_iter,
scan_part_labels,
smplx_part_labels,
search_tree=None,
pen_distance=None,
tri_filtering_module=None,
display=None):
"""
Optimize SMPLX.
:param display: if not None, pass index of the scan in th_scan_meshes to visualize.
"""
# smplx.expression.requires_grad = False
# smplx.jaw_pose.requires_grad = False
# Optimizer
optimizer = torch.optim.Adam([
smplx.trans, smplx.betas, smplx.global_pose, smplx.body_pose,
smplx.left_hand_pose, smplx.right_hand_pose
],
0.02,
betas=(0.9, 0.999))
# Get loss_weights
weight_dict = get_loss_weights()
# Display
if display is not None:
assert int(display) < len(th_scan_meshes)
mv = MeshViewer()
for it in range(iterations):
loop = tqdm(range(steps_per_iter))
loop.set_description('Optimizing SMPLX')
for i in loop:
optimizer.zero_grad()
# Get losses for a forward pass
loss_dict = forward_step(th_scan_meshes, smplx, scan_part_labels,
smplx_part_labels, search_tree,
pen_distance, tri_filtering_module)
# Get total loss for backward pass
tot_loss = backward_step(loss_dict, weight_dict, it)
tot_loss.backward()
optimizer.step()
l_str = 'Iter: {}'.format(i)
for k in loss_dict:
l_str += ', {}: {:0.4f}'.format(
k, weight_dict[k](loss_dict[k], it).mean().item())
loop.set_description(l_str)
if display is not None:
# verts, _, _, _ = smplx()
verts = smplx()
smplx_mesh = Mesh(v=verts[display].cpu().detach().numpy(),
f=smplx.faces.cpu().numpy())
scan_mesh = Mesh(
v=th_scan_meshes[display].vertices.cpu().detach().numpy(),
f=th_scan_meshes[display].faces.cpu().numpy(),
vc=np.array([0, 1, 0]))
scan_mesh.set_vertex_colors_from_weights(
scan_part_labels[display].cpu().detach().numpy())
mv.set_static_meshes([scan_mesh, smplx_mesh])
print('** Optimised smplx pose and shape **')
def optimize_pose_only(th_scan_meshes,
smpl,
iterations,
steps_per_iter,
th_pose_3d,
prior_weight,
display=None):
"""
Initially we want to only optimize the global rotation of SMPL. Next we optimize full pose.
We optimize pose based on the 3D keypoints in th_pose_3d.
:param th_pose_3d: array containing the 3D keypoints.
:param prior_weight: weights corresponding to joints depending on visibility of the joint in the 3D scan.
eg: hand could be inside pocket.
"""
batch_sz = smpl.pose.shape[0]
split_smpl = th_batch_SMPL_split_params(batch_sz,
top_betas=smpl.betas.data[:, :2],
other_betas=smpl.betas.data[:, 2:],
global_pose=smpl.pose.data[:, :3],
other_pose=smpl.pose.data[:, 3:],
faces=smpl.faces,
gender=smpl.gender).cuda()
optimizer = torch.optim.Adam(
[split_smpl.trans, split_smpl.top_betas, split_smpl.global_pose],
0.02,
betas=(0.9, 0.999))
# Get loss_weights
weight_dict = get_loss_weights()
if display is not None:
assert int(display) < len(th_scan_meshes)
# mvs = MeshViewers((1,1))
mv = MeshViewer(keepalive=True)
iter_for_global = 1
for it in range(iter_for_global + iterations):
loop = tqdm(range(steps_per_iter))
if it < iter_for_global:
# Optimize global orientation
print('Optimizing SMPL global orientation')
loop.set_description('Optimizing SMPL global orientation')
elif it == iter_for_global:
# Now optimize full SMPL pose
print('Optimizing SMPL pose only')
loop.set_description('Optimizing SMPL pose only')
optimizer = torch.optim.Adam([
split_smpl.trans, split_smpl.top_betas, split_smpl.global_pose,
split_smpl.other_pose
],
0.02,
betas=(0.9, 0.999))
else:
loop.set_description('Optimizing SMPL pose only')
for i in loop:
optimizer.zero_grad()
# Get losses for a forward pass
loss_dict = forward_step_pose_only(split_smpl, th_pose_3d,
prior_weight)
# Get total loss for backward pass
tot_loss = backward_step(loss_dict, weight_dict, it)
tot_loss.backward()
optimizer.step()
l_str = 'Iter: {}'.format(i)
for k in loss_dict:
l_str += ', {}: {:0.4f}'.format(
k, weight_dict[k](loss_dict[k], it).mean().item())
loop.set_description(l_str)
if display is not None:
verts, _, _, _ = split_smpl()
smpl_mesh = Mesh(v=verts[display].cpu().detach().numpy(),
f=smpl.faces.cpu().numpy())
scan_mesh = Mesh(
v=th_scan_meshes[display].vertices.cpu().detach().numpy(),
f=th_scan_meshes[display].faces.cpu().numpy(),
vc=np.array([0, 1, 0]))
mv.set_dynamic_meshes([smpl_mesh, scan_mesh])
# from matplotlib import cm
# col = cm.tab20c(np.arange(len(th_pose_3d[display]['pose_keypoints_3d'])) % 20)[:, :3]
#
# jts, _, _ = split_smpl.get_landmarks()
# Js = plot_points(jts[display].detach().cpu().numpy(), cols=col)
# Js_observed = plot_points(th_pose_3d[display]['pose_keypoints_3d'][:, :3].numpy(), cols=col)
# mvs[0][0].set_static_meshes([smpl_mesh, scan_mesh])
# mvs[0][1].set_static_meshes(Js)
# mvs[0][2].set_static_meshes(Js_observed)
# Put back pose, shape and trans into original smpl
smpl.pose.data = split_smpl.pose.data
smpl.betas.data = split_smpl.betas.data
smpl.trans.data = split_smpl.trans.data
print('** Optimised smpl pose **')
if mesh.v.max() > 1e3:
mesh.v /= 1e3
n_vertices = len(mesh.v)
prediction_vertex_values = np.arange(n_vertices)
minima = min(prediction_vertex_values)
maxima = max(prediction_vertex_values)
norm = matplotlib.colors.Normalize(vmin=minima, vmax=maxima, clip=True)
mapper = cm.ScalarMappable(norm=norm, cmap=cm.viridis)
mapper_list = [mapper.to_rgba(x)[0] for x in prediction_vertex_values]
mesh.vc = mapper_list
mv = MeshViewer()
mv.dynamic_meshes = [mesh]
time.sleep(100)
mv.save_snapshot(
'/home/eman/Documents/PhD/pytorch_geometric-master/examples/reference-mesh.png'
)
####################################################################
if args.case == 'MPI':
data_path = '/home/eman/Documents/PhD/body-modeling-master_initial/smpl-viewer/Data-survey-smpl/AllData/'
dirs = os.listdir(data_path)
dirs.sort()
print(dirs)
print len(objFiles)
for i in range(len(objFiles)):
#Displaying original faces
#i = 24
originalMeshFile = Mesh(filename=path + objFiles[i])
originalMeshFile.texture_filepath = path + bmpFiles[i]
leftVerts = originalMeshFile
rightVerts = originalMeshFile
#print vars(leftVerts)
name = (objFiles[i])[:-4]
mv = MeshViewer()
mv.dynamic_meshes = [originalMeshFile]
time.sleep(0.3)
mv.save_snapshot(saving_path + name + '_frontal.png')
#time.sleep(0.1)
#mv.
#Left view
newleftVerts = (np.dot(rotation_matrix(axis, theta_l),
(leftVerts.v).transpose()))
leftVerts.v = newleftVerts.transpose()
#mv_l = MeshViewer()
mv.dynamic_meshes = [leftVerts]
time.sleep(0.3)
mv.save_snapshot(saving_path + name + '_left.png')
