def to_ply(self, filepath):
"""
Write mesh to OBJ file.
:param filepath: path to OBJ file
:type filepath: str
"""
utils.write_ply(filepath, self.vertices.tolist(), self.faces.tolist())
## For loop over the input images
count = 0
listImgs = glob("tmp_ims/*.png")
for image_path in listImgs:
start_each_image = time.time()
count = count + 1
fig_name = ntpath.basename(image_path)
outFile = data_out + "/" + fig_name[:-4]
print '> Processing image: ', image_path, ' ', fig_name, ' ', str(count) + '/' + str(len(listImgs))
net.blobs['data'].reshape(1,3,trg_size,trg_size)
im = caffe.io.load_image(image_path)
## Transforming the image into the right format
net.blobs['data'].data[...] = transformer.preprocess('data', im)
## Forward pass into the CNN
net_output = net.forward()
## Getting the output
features = np.hstack( [net.blobs[layer_name].data[0].flatten()] )
## Writing the regressed 3DMM parameters
np.savetxt(outFile + '.ply.alpha', features[0:99])
np.savetxt(outFile + '.ply.beta', features[99:198])
#################################
## Mapping back the regressed 3DMM into the original
## Basel Face Model (Shape)
##################################
S,T = utils.projectBackBFM(model,features)
print '> Writing 3D file in: ', outFile + '.ply' + 'cost {0}s.'.format(time.time()-start_each_image)
utils.write_ply(outFile + '.ply', S, T, faces)
end_time = time.time()
print "Costtime{0}s".format(end_time - start_time)
import os
import numpy as np
import tqdm
import utils
PLY_DIR = "dataset/ply_files"
if __name__ == "__main__":
if not os.path.exists(PLY_DIR):
os.mkdir(PLY_DIR)
# Get all npy files paths
files = glob.glob("{}/**/*.npy".format(utils.NPY_DIR), recursive=True)
# For each file, generate its ply file
for f in tqdm.tqdm(files):
data = np.load(f)
# File name without extension
file_name = f.split("/")[-1][:-4]
utils.write_ply(
"{}/{}.ply".format(PLY_DIR, file_name),
[data[:, 0], data[:, 1], data[:, 2],
data[:, 3].astype(np.uint8), data[:, 4].astype(np.uint8),
data[:, 5].astype(np.uint8), data[:, 6].astype(np.int32)],
["x", "y", "z", "red", "green", "blue", "label"],
)
for voxel_center_index in progress_bar:
neighbors_filter = neighbors_indexes[voxel_center_index]
points_in_voxel = point_cloud[neighbors_filter]
if points_in_voxel.shape[0] > 0:
matrix = voxel_values[voxel_center_index -
nb_voxel_in_voxel_grid //
2:voxel_center_index +
nb_voxel_in_voxel_grid // 2]
matrix = np.reshape(matrix, (1, 1, args.N, args.N, args.N))
out = model(torch.from_numpy(matrix).float())
label = torch.argmax(out)
predicted_labels[neighbors_filter] = label
# Create ply
# ---------------
utils.write_ply(
"{}/{}.ply".format(RESULTS_DIR,
args.data.split("/")[-1].split(".")[-2]),
[
point_cloud[:, 0], point_cloud[:, 1], point_cloud[:, 2],
predicted_labels.astype(np.int32), point_cloud[:, 6].astype(
np.int32)
],
["x", "y", "z", "predicted_labels", "true_labels"],
)
def run():
###################################################
##### Prepare images ##############################
countIms = 0
with open(fileList, "r") as ins, open(data_out + "/imList.txt","w") as outs:
for image_path in ins:
# image_path = image_path[:-1]
print("> Prepare image "+image_path + ":")
imname = ntpath.basename(image_path)
#imname = imname[:-4]
imname = imname.split(imname.split('.')[-1])[0][0:-1]
img = cv2.imread(image_path)
## If we have input landmarks
if len(landmarkDir) > 0:
lms = np.loadtxt(landmarkDir + '/' + imname + '.pts')
img2 = cv2.copyMakeBorder(img,0,0,0,0,cv2.BORDER_REPLICATE)
nLM = lms.shape[0]
for i in range(0,nLM):
cv2.circle(img2, (lms[i,0], lms[i,1]), 5, (255,0,0))
img, lms = utils.cropByInputLM(img, lms, img2)
else:
dlib_img = io.imread(image_path)
img2 = cv2.copyMakeBorder(img,0,0,0,0,cv2.BORDER_REPLICATE)
dets = detector(img, 1)
print("> Number of faces detected: {}".format(len(dets)))
if len(dets) == 0:
print('> Could not detect the face, skipping the image...' + image_path)
continue
if len(dets) > 1:
print("> Process only the first detected face!")
detected_face = dets[0]
## If we are using landmarks to crop
shape = predictor(dlib_img, detected_face)
nLM = shape.num_parts
for i in range(0,nLM):
cv2.circle(img2, (shape.part(i).x, shape.part(i).y), 5, (255,0,0))
img, lms = utils.cropByLM(img, shape, img2)
cv2.imwrite(data_out + "/imgs/"+imname+"_detect.png",img2)
lms = lms * 500.0/img.shape[0]
fileout = open(data_out + "/imgs/"+imname + ".pts","w")
for i in range(0,lms.shape[0]):
fileout.write("%f %f\n" % (lms[i,0], lms[i,1]))
fileout.close()
img = cv2.resize(img,(500, 500))
cv2.imwrite(data_out + "/imgs/"+imname+ ".png",img)
outs.write("%s\n" % (data_out + "/imgs/"+imname+ ".png"))
countIms = countIms + 1
###################################################
##### Shape fitting ##############################
# load net
MainModel = imp.load_source('MainModel', "../CNN/shape_model.py")
net = torch.load(model_path)
net.eval()
mean0 = np.load(mean_path, encoding='latin1')
mean = mean0['arr_0']
net.cuda()
print('> CNN Model loaded to regress 3D Shape and Texture!')
model = scipy.io.loadmat(BFM_path,squeeze_me=True,struct_as_record=False)
model = model["BFM"]
faces = model.faces-1
print('> Loaded the Basel Face Model to write the 3D output!')
## For loop over the input images
count = 0
with open(data_out + "/imList.txt", "r") as ins:
for image_path in ins:
if len(image_path) < 3:
continue
image_path = image_path[:-1]
count = count + 1
fig_name = ntpath.basename(image_path)
outFile = data_out + "/shape/" + fig_name[:-4]
print('> Processing image: ' + image_path)
im = cv2.imread(image_path)
im = cv2.resize(im, (224, 224)).astype(float).transpose((2,0,1))
im = im - mean
#im = im/255
im = Variable(torch.from_numpy(im).unsqueeze(0).float().cuda())
features = net(im).data.cpu().numpy()
## Writing the regressed 3DMM parameters
np.savetxt(outFile + '.ply.alpha', features[0,0:99])
S,T = utils.projectBackBFM(model,features[0,:])
print('> Writing 3D file in: ', outFile + '.ply')
utils.write_ply(outFile + '.ply', S, T, faces)
##################################################
##### Bump map regression ########################
print("Regress bump maps")
bumpMapRegressor.estimateBump(bumpModel_path, data_out + "/imList.txt", data_out + "/bump/")
##################################################
##### Recover the 3D models ##################
print("Recover the 3D models")
print("./TestBump -batch " + data_out + "/imList.txt " + data_out + "/3D/ " + data_out + "/shape " + data_out + "/bump " + data_out + "/bump ../3DMM_model/BaselFaceModel_mod.h5 ../dlib_model/shape_predictor_68_face_landmarks.dat " + data_out + "/imgs " + data_out + "/imgs/ 1");
os.system("./TestBump -batch " + data_out + "/imList.txt " + data_out + "/3D/ " + data_out + "/shape " + data_out + "/bump " + data_out + "/bump ../3DMM_model/BaselFaceModel_mod.h5 ../dlib_model/shape_predictor_68_face_landmarks.dat " + data_out + "/imgs " + data_out + "/imgs/ 1");
def process(self):
for path in self.processed_paths:
os.makedirs(path)
for pc_path in glob.glob(os.path.join(self.raw_paths[0], '*.txt')):
print('Processing {} ...'.format(pc_path))
cloud_name = pc_path.split('/')[-1][:-4]
if os.path.exists(os.path.join(self.processed_paths[1], cloud_name + '_KDTree.pkl')):
continue
pc = self._load_cloud(pc_path)
label_path = pc_path[:-4] + '.labels'
if os.path.exists(label_path):
labels = self._load_label(label_path)
org_ply_path = os.path.join(self.processed_paths[0], cloud_name + '.ply')
# Subsample the training set cloud to the same resolution 0.01 as the test set
xyz, rgb, labels = self._grid_sub_sampling(pc[:, :3].astype(np.float32),
pc[:, 4:7].astype(np.uint8),
labels, grid_size=0.01)
labels = np.squeeze(labels)
# save sub-sampled original cloud
write_ply(org_ply_path, [xyz, rgb, labels], ['x', 'y', 'z', 'r', 'g', 'b', 'class'])
# save sub_cloud and KDTree file
sub_xyz, sub_rgb, sub_labels = self._grid_sub_sampling(xyz, rgb, labels, grid_size=self.grid_size)
sub_rgb = sub_rgb / 255.
sub_labels = np.squeeze(sub_labels)
sub_ply_file = os.path.join(self.processed_paths[1], cloud_name + '.ply')
write_ply(sub_ply_file, [sub_xyz, sub_rgb, sub_labels], ['x', 'y', 'z', 'r', 'g', 'b', 'class'])
search_tree = KDTree(sub_xyz, leaf_size=50)
kd_tree_file = os.path.join(self.processed_paths[1], cloud_name + '_KDTree.pkl')
with open(kd_tree_file, 'wb') as f:
pickle.dump(search_tree, f)
proj_idx = np.squeeze(search_tree.query(xyz, return_distance=False))
proj_idx = proj_idx.astype(np.int32)
proj_file = os.path.join(self.processed_paths[1], cloud_name + '_proj.pkl')
with open(proj_file, 'wb') as f:
pickle.dump([proj_idx, labels], f)
else:
org_ply_path = os.path.join(self.processed_paths[0], cloud_name + '.ply')
write_ply(org_ply_path, [pc[:, :3].astype(np.float32), pc[:, 4:7].astype(np.uint8)],
['x', 'y', 'z', 'r', 'g', 'b'])
sub_xyz, sub_rgb = self._grid_sub_sampling(pc[:, :3].astype(np.float32),
pc[:, 4:7].astype(np.uint8),
grid_size=self.grid_size)
sub_rgb = sub_rgb / 255.
sub_ply_file = os.path.join(self.processed_paths[1], cloud_name + '.ply')
write_ply(sub_ply_file, [sub_xyz, sub_rgb], ['x', 'y', 'z', 'r', 'g', 'b'])
search_tree = KDTree(sub_xyz, leaf_size=50)
kd_tree_file = os.path.join(self.processed_paths[1], cloud_name + '_KDTree.pkl')
with open(kd_tree_file, 'wb') as f:
pickle.dump(search_tree, f)
labels = np.zeros(pc.shape[0], dtype=np.uint8)
proj_idx = np.squeeze(search_tree.query(pc[:, :3].astype(np.float32), return_distance=False))
proj_idx = proj_idx.astype(np.int32)
proj_file = os.path.join(self.processed_paths[1], cloud_name + '_proj.pkl')
with open(proj_file, 'wb') as f:
pickle.dump([proj_idx, labels], f)
def cnn_3dmm( img, outputPath ):
"""
use 3dmm cnn to deal with
"""
# load net
try:
caffe.set_mode_gpu()
caffe.set_device(GPU_ID)
except Exception as ex:
print '> Could not setup Caffe in GPU ' +str(GPU_ID) + ' - Error: ' + ex
print '> Reverting into CPU mode'
caffe.set_mode_cpu()
## Opening mean average image
proto_data = open(FLAGS.mean_path, "rb").read()
a = caffe.io.caffe_pb2.BlobProto.FromString(proto_data)
mean = caffe.io.blobproto_to_array(a)[0]
## Loading the CNN
net = caffe.Classifier(FLAGS.deploy_path, FLAGS.model_path)
## Setting up the right transformer for an input image
transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})
transformer.set_transpose('data', (2,0,1))
transformer.set_channel_swap('data', (2,1,0))
transformer.set_raw_scale('data', 255.0)
transformer.set_mean('data',mean)
print '> CNN Model loaded to regress 3D Shape and Texture!'
## Loading the Basel Face Model to write the 3D output
model = scipy.io.loadmat(FLAGS.BFM_path,squeeze_me=True,struct_as_record=False)
model = model["BFM"]
faces = model.faces-1
print '> Loaded the Basel Face Model to write the 3D output!'
net.blobs['data'].reshape(1,3,trg_size,trg_size)
# img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
# im = img / 255
imname = ntpath.basename(FLAGS.imagePath)
imname = imname.split(imname.split('.')[-1])[0][0:-1] # 记录图片basename
im = caffe.io.load_image(os.path.join(FLAGS.tmp_ims, imname + '.png'))
## Transforming the image into the right format
net.blobs['data'].data[...] = transformer.preprocess('data', im)
## Forward pass into the CNN
net_output = net.forward()
## Getting the output
features = np.hstack( [net.blobs[layer_name].data[0].flatten()] )
#imname = ntpath.basename(FLAGS.imagePath)
#imname = imname.split(imname.split('.')[-1])[0][0:-1] # 记录图片basename
## Writing the regressed 3DMM parameters
np.savetxt(os.path.join( outputPath, imname+ '.ply.alpha'), features[0:99])
np.savetxt(os.path.join( outputPath, imname + '.ply.beta'), features[99:198])
#################################
## Mapping back the regressed 3DMM into the original
## Basel Face Model (Shape)
##################################
S,T = utils.projectBackBFM(model, features)
print '> Writing 3D file in: ', os.path.join( outputPath, imname+ '.ply')
utils.write_ply(os.path.join( outputPath, imname+ '.ply'), S, T, faces)
verts = pc_np[mask_np.reshape(-1), :]
normal = normal.cpu().squeeze(0).view(
3, -1).t().numpy()[mask_np.reshape(-1), :]
props = [
verts[:, 0], verts[:, 1], verts[:, 2], normal[:, 0], normal[:, 1],
normal[:, 2]
]
prop_names = ['x', 'y', 'z', 'nx', 'ny', 'nz']
prop_types = ['float32' for _ in range(0, 6)]
#has albedo
if albedo_path is not None:
albedo = albedo_np.reshape(-1, 3)[mask_np.reshape(-1), :]
props += [albedo[:, 0], albedo[:, 1], albedo[:, 2]]
prop_names += ['albedo_r', 'albedo_g', 'albedo_b']
prop_types += ['float32' for _ in range(0, 3)]
#has rough
if rough_path is not None:
rough = rough_np.reshape(-1)[mask_np.reshape(-1)]
props += [rough]
prop_names += ['rough_r']
prop_types += ['float32']
write_ply(output_path,
props,
prop_names=prop_names,
prop_types=prop_types,
faces=faces)
def test(self, num_votes=100):
logging.info('Test {} on {} ...'.format(self.cfg.model_name,
self.cfg.dataset))
test_smooth = 0.98
saving_path = 'results/Semantic3D/predictions'
os.makedirs(saving_path) if not os.path.exists(saving_path) else None
# load model checkpoints
self.model.load(
'checkpoints/PointConvBig_on_Semantic3D_bs_8_epochs_100_big_crf.ckpt'
)
self.model.to(self.device)
self.model.eval()
epoch = 0
last_min = -0.5
while last_min < num_votes:
# test one epoch
with Ctq(self.dataset.val_loader) as tq_loader:
for i, data in enumerate(tq_loader):
tq_loader.set_description('Evaluation')
# model inference
data = data.to(self.device)
with torch.no_grad():
probs = F.softmax(self.model(data),
dim=-1) # get pred probs
# running means for each epoch on Test set
point_idx = data.point_idx.cpu().numpy() # the point idx
cloud_idx = data.cloud_idx.cpu().numpy() # the cloud idx
probs = probs.reshape(
self.cfg.batch_size, -1,
self.cfg.num_classes).cpu().numpy() # [B, N, C]
for b in range(
self.cfg.batch_size): # for each sample in batch
prob = probs[b, :, :] # [N, C]
p_idx = point_idx[b, :] # [N]
c_idx = cloud_idx[b][0] # int
self.test_probs[c_idx][p_idx] = test_smooth * self.test_probs[c_idx][p_idx] \
+ (1 - test_smooth) * prob # running means
# after each epoch
new_min = np.min(self.dataset.val_set.min_possibility)
print('Epoch {:3d} end, current min possibility = {:.2f}'.format(
epoch, new_min))
if last_min + 4 < new_min:
print('Test procedure done, saving predicted clouds ...')
last_min = new_min
# projection prediction to original point cloud
t1 = time.time()
for i, file in enumerate(self.dataset.val_set.val_files):
proj_idx = self.dataset.val_set.test_proj[
i] # already get the shape
probs = self.test_probs[i][
proj_idx, :] # same shape with proj_idx
# [0 ~ 7] + 1 -> [1 ~ 8], because 0 for unlabeled
preds = np.argmax(probs, axis=1).astype(np.uint8) + 1
# saving prediction results
cloud_name = file.split('/')[-1]
# ascii_name = os.path.join(saving_path, self.dataset.test_set.ascii_files[cloud_name])
# np.savetxt(ascii_name, preds, fmt='%d')
# print('Save {:s} succeed !'.format(ascii_name))
filename = os.path.join(saving_path, cloud_name)
write_ply(filename, [preds], ['pred'])
print('Save {:s} succeed !'.format(filename))
t2 = time.time()
print('Done in {:.2f} s.'.format(t2 - t1))
return
epoch += 1
return
def process(self):
# Note: there is an extra character in line 180389 of Area_5/hallway_6/Annotations/ceiling_1.txt
for path in self.processed_paths:
os.makedirs(path)
for i, path in enumerate(self.raw_paths):
print("Processing Area_{}...".format(i + 1))
anno_paths = [line.rstrip() for line in open(path)]
anno_paths = [
os.path.join(self.raw_dir, self.data_dir, p)
for p in anno_paths
]
for anno_path in anno_paths:
print('Processing {}...'.format(anno_path))
elements = anno_path.split('/')
filename = elements[-3] + '_' + elements[-2]
data_list = []
for f in glob.glob(os.path.join(anno_path, '*.txt')):
print('Collecting {}...'.format(f))
label = os.path.basename(f).split('_')[0]
if label not in CLASS_NAMES:
label = 'clutter'
# cls_points = np.loadtxt(f)
cls_points = pd.read_csv(
f, header=None, delim_whitespace=True
).values # pandas for faster reading
cls_labels = np.full((cls_points.shape[0], 1),
CLASS_NAMES[label],
dtype=np.int32)
data_list.append(
np.concatenate([cls_points, cls_labels],
axis=1)) # Nx7
points_labels = np.concatenate(data_list, axis=0)
xyz_min = np.amin(points_labels, axis=0)[0:3]
points_labels[:,
0:3] -= xyz_min # aligned to the minimal point
xyz = points_labels[:, 0:3].astype(np.float32)
rgb = points_labels[:, 3:6].astype(np.uint8)
labels = points_labels[:, 6].astype(np.uint8)
org_ply_file = os.path.join(self.processed_paths[0],
filename + '.ply')
write_ply(org_ply_file, [xyz, rgb, labels],
['x', 'y', 'z', 'r', 'g', 'b', 'class'])
# save sub_cloud and KDTree files
sub_xyz, sub_rgb, sub_labels = self._grid_sub_sampling(
xyz, rgb, labels, self.grid_size)
sub_rgb = sub_rgb / 255.
sub_ply_file = os.path.join(self.processed_paths[1],
filename + '.ply')
write_ply(sub_ply_file, [sub_xyz, sub_rgb, sub_labels],
['x', 'y', 'z', 'r', 'g', 'b', 'class'])
search_tree = KDTree(sub_xyz)
kd_tree_file = os.path.join(self.processed_paths[1],
filename + '_KDTree.pkl')
with open(kd_tree_file, 'wb') as f:
pickle.dump(search_tree, f)
proj_idx = np.squeeze(
search_tree.query(xyz, return_distance=False))
proj_idx = proj_idx.astype(np.int32)
proj_file = os.path.join(self.processed_paths[1],
filename + '_proj.pkl')
with open(proj_file, 'wb') as f:
pickle.dump([proj_idx, labels], f)
