def __init__(self,
num_scales=4,
ssim_loss_weight=0.85,
occ_reg_weight=0.1,
smooth_loss_weight=0.1,
C1=1e-4,
C2=9e-4,
photometric_reduce_op='mean',
disp_norm=True,
clip_loss=0.5,
progressive_scaling=0.0,
padding_mode='zeros',
automask_loss=False,
**kwargs):
super().__init__()
self.n = num_scales
self.progressive_scaling = progressive_scaling
self.ssim_loss_weight = ssim_loss_weight
self.occ_reg_weight = occ_reg_weight
self.smooth_loss_weight = smooth_loss_weight
self.C1 = C1
self.C2 = C2
self.photometric_reduce_op = photometric_reduce_op
self.disp_norm = disp_norm
self.clip_loss = clip_loss
self.padding_mode = padding_mode
self.automask_loss = automask_loss
self.progressive_scaling = ProgressiveScaling(progressive_scaling,
self.n)
# Asserts
if self.automask_loss:
assert self.photometric_reduce_op == 'min', \
'For automasking only the min photometric_reduce_op is supported.'
def __init__(self, num_scales=1, ssim_loss_weight=0.85, occ_reg_weight=0.1, smooth_loss_weight=0.1,
C1=1e-4, C2=9e-4, photometric_reduce_op='mean', disp_norm=True, clip_loss=0.5,
progressive_scaling=0.0, padding_mode='zeros',
automask_loss=False, **kwargs):
super().__init__()
self.n = num_scales
self.progressive_scaling = progressive_scaling
self.ssim_loss_weight = ssim_loss_weight
self.occ_reg_weight = occ_reg_weight
self.smooth_loss_weight = smooth_loss_weight
self.C1 = C1
self.C2 = C2
self.photometric_reduce_op = photometric_reduce_op
self.disp_norm = disp_norm
self.clip_loss = clip_loss
self.padding_mode = padding_mode
self.automask_loss = automask_loss
self.progressive_scaling = ProgressiveScaling(
progressive_scaling, self.n)
self.canonical_ray_surface = torch.tensor(
#np.load("kitti_ray_template.npy"))
#np.load("omnicam_ray_template.npy"))
#np.load("dummy_ray_template_240_384.npy"))
#np.load("dummy_ray_template_224_384.npy"))
np.load("dummy_ray_template_512_320.npy"))
# Asserts
if self.automask_loss:
assert self.photometric_reduce_op == 'min', \
'For automasking only the min photometric_reduce_op is supported.'
def __init__(self, num_scales=4,
progressive_scaling=0.0, mask_ego=True, mask_out_of_bounds_reprojected=True, **kwargs):
super().__init__()
self.n = num_scales
self.progressive_scaling = progressive_scaling
self.mask_ego = mask_ego
self.mask_out_of_bounds_reprojected = mask_out_of_bounds_reprojected
self.progressive_scaling = ProgressiveScaling(
progressive_scaling, self.n)
def __init__(self,
supervised_method='sparse-l1',
supervised_num_scales=4,
progressive_scaling=0.0,
**kwargs):
super().__init__()
self.loss_func = get_loss_func(supervised_method)
self.supervised_method = supervised_method
self.n = supervised_num_scales
self.progressive_scaling = ProgressiveScaling(progressive_scaling,
self.n)
def __init__(self,
num_scales=4,
ssim_loss_weight=0.85,
occ_reg_weight=0.1,
smooth_loss_weight=0.1,
depth_consistency_loss_weight=0.1,
scene_flow_consistency_loss_weight=0.1,
C1=1e-4,
C2=9e-4,
photometric_reduce_op='mean',
disp_norm=True,
sort_cameras=False,
sort_swap=[],
clip_loss=0.5,
progressive_scaling=0.0,
padding_mode='zeros',
temporal_loss_weight=1.0,
spatial_loss_weight=0.1,
temporal_spatial_loss_weight=0.1,
automask_loss=True,
cameras=None,
**kwargs):
super().__init__()
self.n = num_scales
self.progressive_scaling = progressive_scaling
self.ssim_loss_weight = ssim_loss_weight
self.occ_reg_weight = occ_reg_weight
self.smooth_loss_weight = smooth_loss_weight
self.depth_consistency_loss_weight = depth_consistency_loss_weight
self.scene_flow_consistency_loss_weight = scene_flow_consistency_loss_weight
self.C1 = C1
self.C2 = C2
self.photometric_reduce_op = photometric_reduce_op
self.disp_norm = disp_norm
self.clip_loss = clip_loss
self.padding_mode = padding_mode
self.automask_loss = automask_loss
self.temporal_loss_weight = temporal_loss_weight
self.spatial_loss_weight = spatial_loss_weight
self.temporal_spatial_loss_weight = temporal_spatial_loss_weight
self.progressive_scaling = ProgressiveScaling(progressive_scaling,
self.n)
self.sort_cameras = sort_cameras
self.sort_swap = sort_swap
self.cameras = cameras
# Asserts
if self.automask_loss:
assert self.photometric_reduce_op == 'min', \
'For automasking only the min photometric_reduce_op is supported.'
def __init__(self, num_scales=4, ssim_loss_weight=0.85, occ_reg_weight=0.1, smooth_loss_weight=0.1,
C1=1e-4, C2=9e-4, photometric_reduce_op='min', disp_norm=True, clip_loss=0.5,
progressive_scaling=0.0, padding_mode='zeros',
automask_loss=False,
mask_ego=True,
mask_occlusion=False, mask_disocclusion=False, mask_spatial_context=False, mask_temporal_context=False,
depth_consistency_weight=0.2,
mult_margin_occlusion=1.5, add_margin_occlusion=1.5,
allow_context_rotation=False,
**kwargs):
super().__init__()
self.n = num_scales
self.progressive_scaling = progressive_scaling
self.ssim_loss_weight = ssim_loss_weight
self.occ_reg_weight = occ_reg_weight
self.smooth_loss_weight = smooth_loss_weight
self.C1 = C1
self.C2 = C2
self.photometric_reduce_op = photometric_reduce_op
self.disp_norm = disp_norm
self.clip_loss = clip_loss
self.padding_mode = padding_mode
self.automask_loss = automask_loss
self.mask_ego = mask_ego
self.mask_occlusion = mask_occlusion
self.mask_disocclusion = mask_disocclusion
self.mask_spatial_context = mask_spatial_context
self.mask_temporal_context = mask_temporal_context
self.mult_margin_occlusion = mult_margin_occlusion
self.add_margin_occlusion = add_margin_occlusion
self.depth_consistency_weight = depth_consistency_weight
self.allow_context_rotation = allow_context_rotation
self.use_ref_depth = ((self.mask_occlusion or self.mask_disocclusion) and (self.mask_spatial_context or self.mask_temporal_context)) or (self.depth_consistency_weight > 0)
self.progressive_scaling = ProgressiveScaling(
progressive_scaling, self.n)
# Asserts
if self.automask_loss:
assert self.photometric_reduce_op == 'min', \
'For automasking only the min photometric_reduce_op is supported.'