depth_mapping_3d = np.ones(vox_size[2] * vox_size[1] * vox_size[0],
dtype=np.float32) * (-1)
dp.TSDF((depth / 1000.0).reshape(-1), cam_K, cam_pose, vox_origin,
vox_unit, vox_size.astype(np.float32), vox_margin, depth.shape[0],
depth.shape[1], tsdf, depth_mapping_3d)
depth_mapping_3d = depth_mapping_3d.astype(np.int64)
label = np.zeros(vox_size[2] * vox_size[1] * vox_size[0] /
(sampleRatio * sampleRatio * sampleRatio),
dtype=np.float32)
tsdf_downsample = np.zeros(vox_size[2] * vox_size[1] * vox_size[0] /
(sampleRatio * sampleRatio * sampleRatio),
dtype=np.float32)
dp.DownSampleLabel(readin_bin, vox_size.astype(np.float32), sampleRatio,
tsdf, label, tsdf_downsample)
label = label.astype(np.int32)
label_weight = getLabelWeight(label, tsdf_downsample)
label[np.where(label == 255)] = 0
dp.TSDFTransform(tsdf, vox_size.astype(np.float32))
tsdf = torch.FloatTensor(tsdf).view(
(1, vox_size[2], vox_size[1], vox_size[0]))
label = torch.LongTensor(label).view(vox_size[2] / sampleRatio,
vox_size[1] / sampleRatio,
vox_size[0] / sampleRatio)
label_weight = torch.FloatTensor(label_weight).view(
vox_size[2] / sampleRatio, vox_size[1] / sampleRatio,
vox_size[0] / sampleRatio)
depth_mapping_3d = torch.LongTensor(depth_mapping_3d)
np.savez_compressed(os.path.join(outputpath, '%06d.npz' % fileindex),
tsdf.numpy(), label.numpy(), label_weight.numpy(),
def multi_process(index, lenlist, startindex):
print str(index) + '/' + str(lenlist)
binfile = open(binlist[index - startindex], 'rb')
vox_origin = np.array(struct.unpack('3f', binfile.read(12)),
dtype=np.float32)
cam_pose = np.array(struct.unpack('16f', binfile.read(64)),
dtype=np.float32)
readin_bin = np.zeros(vox_size[0] * vox_size[1] * vox_size[2],
dtype=np.float32)
all_the_byte = binfile.read()
binfile.close()
all_the_val = struct.unpack("%dI" % (len(all_the_byte) / 4), all_the_byte)
last_index = 0
count_bad_data = 0
for i in xrange(len(all_the_val) / 2):
if last_index + all_the_val[
i * 2 + 1] > vox_size[0] * vox_size[1] * vox_size[2]:
count_bad_data = 0
break
if all_the_val[i * 2] == 255:
readin_bin[last_index:last_index + all_the_val[i * 2 + 1]] = 255
else:
readin_bin[last_index:last_index +
all_the_val[i * 2 + 1]] = segmentation_class_map[
all_the_val[i * 2] % 37]
if all_the_val[i * 2] != 0:
count_bad_data += all_the_val[i * 2 + 1]
last_index += all_the_val[i * 2 + 1]
if count_bad_data < 10 or last_index != vox_size[0] * vox_size[
1] * vox_size[2]:
print str(index) + ' is awful!'
return
depth = cv2.imread(depthlist[index - startindex], -1)
depth = ((depth << 13) | (depth >> 3)).astype(np.float32)
tsdf = np.ones(vox_size[2] * vox_size[1] * vox_size[0],
dtype=np.float32) # modified: zeros to ones
depth_mapping_3d = np.ones(640 * 480, dtype=np.float32) * (
-1
) # vox_size[2]*vox_size[1]*vox_size[0] seems to be wrong, it should be 640x480
dp.TSDF((depth / 1000.0).reshape(-1), cam_K, cam_pose, vox_origin,
vox_unit, vox_size.astype(np.float32), vox_margin, depth.shape[0],
depth.shape[1], tsdf, depth_mapping_3d)
depth_mapping_3d = depth_mapping_3d.astype(np.int64)
label = np.zeros(vox_size[2] * vox_size[1] * vox_size[0] /
(sampleRatio * sampleRatio * sampleRatio),
dtype=np.float32)
tsdf_downsample = np.zeros(vox_size[2] * vox_size[1] * vox_size[0] /
(sampleRatio * sampleRatio * sampleRatio),
dtype=np.float32)
dp.DownSampleLabel(readin_bin, vox_size.astype(np.float32), sampleRatio,
tsdf, label, tsdf_downsample)
label = label.astype(np.int32)
label_weight = getLabelWeight(label, tsdf_downsample)
label[np.where(label == 255)] = 0
dp.TSDFTransform(tsdf, vox_size.astype(np.float32))
tsdf = torch.FloatTensor(tsdf).view(
(1, vox_size[2], vox_size[1], vox_size[0]))
label = torch.LongTensor(label).view(vox_size[2] / sampleRatio,
vox_size[1] / sampleRatio,
vox_size[0] / sampleRatio)
label_weight = torch.FloatTensor(label_weight).view(
vox_size[2] / sampleRatio, vox_size[1] / sampleRatio,
vox_size[0] / sampleRatio)
depth_mapping_3d = torch.LongTensor(depth_mapping_3d)
np.savez_compressed(os.path.join(outputpath, '%06d.npz' % index),
tsdf.numpy(), label.numpy(), label_weight.numpy(),
depth_mapping_3d.numpy())
def __getitem__(self, index):
while True:
binfile = open(self.binlist[index], 'rb')
vox_origin = np.array(struct.unpack('3f', binfile.read(12)),
dtype=np.float32)
cam_pose = np.array(struct.unpack('16f', binfile.read(64)),
dtype=np.float32)
readin_bin = np.zeros(self.vox_size[0] * self.vox_size[1] *
self.vox_size[2],
dtype=np.float32)
all_the_byte = binfile.read()
binfile.close()
all_the_val = struct.unpack("%dI" % (len(all_the_byte) / 4),
all_the_byte)
last_index = 0
count_bad_data = 0
for i in xrange(len(all_the_val) / 2):
if last_index + all_the_val[i * 2 + 1] > self.vox_size[
0] * self.vox_size[1] * self.vox_size[2]:
count_bad_data = 0
break
if all_the_val[i * 2] == 255:
readin_bin[last_index:last_index +
all_the_val[i * 2 + 1]] = 255
else:
readin_bin[last_index:last_index +
all_the_val[i * 2 +
1]] = self.segmentation_class_map[
all_the_val[i * 2] % 37]
if all_the_val[i * 2] != 0:
count_bad_data += all_the_val[i * 2 + 1]
last_index += all_the_val[i * 2 + 1]
if count_bad_data < 10 or last_index != self.vox_size[
0] * self.vox_size[1] * self.vox_size[2]:
index = self.hardexamples[random.randint(
0,
len(self.hardexamples) - 1)]
continue
color = cv2.imread(self.colorlist[index], -1)
depth = cv2.imread(self.depthlist[index], -1).astype(np.float32)
tsdf = np.zeros(self.vox_size[2] * self.vox_size[1] *
self.vox_size[0],
dtype=np.float32)
depth_mapping_3d = np.ones(
self.vox_size[2] * self.vox_size[1] * self.vox_size[0],
dtype=np.float32) * (-1)
dp.TSDF((depth / 1000.0).reshape(-1), self.cam_K,
cam_pose, vox_origin, self.vox_unit,
self.vox_size.astype(np.float32), self.vox_margin,
depth.shape[0], depth.shape[1], tsdf, depth_mapping_3d)
depth_mapping_3d = depth_mapping_3d.astype(np.int64)
label = np.zeros(
self.vox_size[2] * self.vox_size[1] * self.vox_size[0] /
(self.sampleRatio * self.sampleRatio * self.sampleRatio),
dtype=np.float32)
tsdf_downsample = np.zeros(
self.vox_size[2] * self.vox_size[1] * self.vox_size[0] /
(self.sampleRatio * self.sampleRatio * self.sampleRatio),
dtype=np.float32)
dp.DownSampleLabel(readin_bin, self.vox_size.astype(np.float32),
self.sampleRatio, tsdf, label, tsdf_downsample)
label = label.astype(np.int32)
label_weight = getLabelWeight(label, tsdf_downsample)
label[np.where(label == 255)] = 0
dp.TSDFTransform(tsdf, self.vox_size.astype(np.float32))
color = np.transpose(color, (2, 0, 1)).astype(np.float32) / 255.0
depth1 = torch.FloatTensor(depth / 10000.0).view(
1, depth.shape[0], depth.shape[1])
color = torch.FloatTensor(color)
color = self.img_transform(color)
img = torch.cat((color, depth1), dim=0).contiguous()
tsdf = torch.FloatTensor(tsdf).view(
(1, self.vox_size[2], self.vox_size[1], self.vox_size[0]))
label = torch.LongTensor(label).view(
self.vox_size[2] / self.sampleRatio,
self.vox_size[1] / self.sampleRatio,
self.vox_size[0] / self.sampleRatio)
label_weight = torch.FloatTensor(label_weight).view(
self.vox_size[2] / self.sampleRatio,
self.vox_size[1] / self.sampleRatio,
self.vox_size[0] / self.sampleRatio)
depth_mapping_3d = torch.LongTensor(depth_mapping_3d)
return img, tsdf, label, label_weight, depth_mapping_3d