def __init__(self, cost_aggregator_inplanes, hourglass_inplanes=16):
super(CostAggregator, self).__init__(inplanes=16)
self.dres0 = nn.Sequential(
convbn_3d_lrelu(cost_aggregator_inplanes, 64, 3, 1, 1),
convbn_3d_lrelu(64, 32, 3, 1, 1))
self.dres1 = nn.Sequential(
convbn_3d_lrelu(32, 32, 3, 1, 1),
convbn_3d_lrelu(32, hourglass_inplanes, 3, 1, 1))
def __init__(self):
super(BGNet_Plus, self).__init__()
self.softmax = nn.Softmax(dim=1)
self.refinement_net = HourglassRefinement()
self.feature_extraction = feature_extraction()
self.coeffs_disparity_predictor = CoeffsPredictor()
self.dres0 = nn.Sequential(convbn_3d_lrelu(44, 32, 3, 1, 1),
convbn_3d_lrelu(32, 16, 3, 1, 1))
self.guide = GuideNN()
self.slice = Slice()
self.weight_init()
def __init__(self, inplanes=16):
super(HourGlass, self).__init__()
self.conv1 = convbn_3d_lrelu(inplanes,
inplanes * 2,
kernel_size=3,
stride=2,
pad=1)
self.conv2 = convbn_3d_lrelu(inplanes * 2,
inplanes * 2,
kernel_size=3,
stride=1,
pad=1)
self.conv1_1 = convbn_3d_lrelu(inplanes * 2,
inplanes * 4,
kernel_size=3,
stride=2,
pad=1)
self.conv2_1 = convbn_3d_lrelu(inplanes * 4,
inplanes * 4,
kernel_size=3,
stride=1,
pad=1)
self.conv3 = convbn_3d_lrelu(inplanes * 4,
inplanes * 8,
kernel_size=3,
stride=2,
pad=1)
self.conv4 = convbn_3d_lrelu(inplanes * 8,
inplanes * 8,
kernel_size=3,
stride=1,
pad=1)
self.conv5 = convbn_transpose_3d(inplanes * 8,
inplanes * 4,
kernel_size=3,
padding=1,
output_padding=(0, 1, 1),
stride=(1, 2, 2),
bias=False)
self.conv6 = convbn_transpose_3d(inplanes * 4,
inplanes * 2,
kernel_size=3,
padding=1,
output_padding=(0, 1, 1),
stride=(1, 2, 2),
bias=False)
self.conv7 = convbn_transpose_3d(inplanes * 2,
inplanes,
kernel_size=3,
padding=1,
output_padding=(0, 1, 1),
stride=(1, 2, 2),
bias=False)
self.last_for_guidance = convbn_3d_lrelu(inplanes,
32,
kernel_size=3,
stride=1,
pad=1)
self.weight_init()
def __init__(self, inplanes=16):
super(HourGlass, self).__init__()
self.conv1 = convbn_3d_lrelu(inplanes,
inplanes * 2,
kernel_size=3,
stride=2,
pad=1)
self.conv2 = convbn_3d_lrelu(inplanes * 2,
inplanes * 2,
kernel_size=3,
stride=1,
pad=1)
self.conv1_1 = convbn_3d_lrelu(inplanes * 2,
inplanes * 4,
kernel_size=3,
stride=2,
pad=1)
self.conv2_1 = convbn_3d_lrelu(inplanes * 4,
inplanes * 4,
kernel_size=3,
stride=1,
pad=1)
self.conv3 = convbn_3d_lrelu(inplanes * 4,
inplanes * 8,
kernel_size=3,
stride=2,
pad=1)
self.conv4 = convbn_3d_lrelu(inplanes * 8,
inplanes * 8,
kernel_size=3,
stride=1,
pad=1)
self.conv5 = convbn_transpose_3d(inplanes * 8,
inplanes * 4,
kernel_size=3,
padding=1,
output_padding=(0, 1, 1),
stride=(1, 2, 2),
bias=False)
self.conv6 = convbn_transpose_3d(inplanes * 4,
inplanes * 2,
kernel_size=3,
padding=1,
output_padding=(0, 1, 1),
stride=(1, 2, 2),
bias=False)
self.conv7 = convbn_transpose_3d(inplanes * 2,
inplanes,
kernel_size=3,
padding=1,
output_padding=(0, 1, 1),
stride=(1, 2, 2),
bias=False)
self.last_conv3d_layer = nn.Sequential(
convbn_3d_lrelu(inplanes, inplanes * 2, 3, 1, 1),
nn.Conv3d(inplanes * 2,
1,
kernel_size=3,
padding=1,
stride=1,
bias=False))
self.softmax = nn.Softmax(dim=1)
self.weight_init()
def __init__(self):
super(DeepPruner, self).__init__()
self.scale = args.cost_aggregator_scale
self.max_disp = args.max_disp // self.scale
self.mode = args.mode
self.patch_match_args = args.patch_match_args
self.patch_match_sample_count = self.patch_match_args.sample_count
self.patch_match_iteration_count = self.patch_match_args.iteration_count
self.patch_match_propagation_filter_size = self.patch_match_args.propagation_filter_size
self.post_CRP_sample_count = args.post_CRP_sample_count
self.post_CRP_sampler_type = args.post_CRP_sampler_type
hourglass_inplanes = args.hourglass_inplanes
# refinement input features are composed of:
# left image low level features +
# CA output features + CA output disparity
if self.scale == 8:
from models.feature_extractor_fast import feature_extraction
refinement_inplanes_1 = args.feature_extractor_refinement_level_1_outplanes + 1
self.refinement_net1 = RefinementNet(refinement_inplanes_1)
else:
from models.feature_extractor_best import feature_extraction
refinement_inplanes = args.feature_extractor_refinement_level_outplanes + self.post_CRP_sample_count + 2 + 1
self.refinement_net = RefinementNet(refinement_inplanes)
# cost_aggregator_inplanes are composed of:
# left and right image features from feature_extractor (ca_level) +
# features from min/max predictors +
# min_disparity + max_disparity + disparity_samples
cost_aggregator_inplanes = 2 * (
args.feature_extractor_ca_level_outplanes +
self.patch_match_sample_count + 2) + 1
self.cost_aggregator = CostAggregator(cost_aggregator_inplanes,
hourglass_inplanes)
self.feature_extraction = feature_extraction()
self.min_disparity_predictor = MinDisparityPredictor(
hourglass_inplanes)
self.max_disparity_predictor = MaxDisparityPredictor(
hourglass_inplanes)
self.spatial_transformer = SpatialTransformer()
self.patch_match = PatchMatch(self.patch_match_propagation_filter_size)
self.uniform_sampler = UniformSampler()
# Confidence Range Predictor(CRP) input features are composed of:
# left and right image features from feature_extractor (ca_level) +
# disparity_samples
CRP_feature_count = 2 * args.feature_extractor_ca_level_outplanes + 1
self.dres0 = nn.Sequential(
convbn_3d_lrelu(CRP_feature_count, 64, 3, 1, 1),
convbn_3d_lrelu(64, 32, 3, 1, 1))
self.dres1 = nn.Sequential(
convbn_3d_lrelu(32, 32, 3, 1, 1),
convbn_3d_lrelu(32, hourglass_inplanes, 3, 1, 1))
self.min_disparity_conv = conv_relu(1, 1, 5, 1, 2)
self.max_disparity_conv = conv_relu(1, 1, 5, 1, 2)
self.ca_disparity_conv = conv_relu(1, 1, 5, 1, 2)
self.ca_features_conv = convbn_2d_lrelu(self.post_CRP_sample_count + 2,
self.post_CRP_sample_count + 2,
5,
1,
2,
dilation=1,
bias=True)
self.min_disparity_features_conv = convbn_2d_lrelu(
self.patch_match_sample_count + 2,
self.patch_match_sample_count + 2,
5,
1,
2,
dilation=1,
bias=True)
self.max_disparity_features_conv = convbn_2d_lrelu(
self.patch_match_sample_count + 2,
self.patch_match_sample_count + 2,
5,
1,
2,
dilation=1,
bias=True)