def forward(self, inputs):
features = inputs['features']
one_hot_vectors = inputs['one_hot_vectors']
assert one_hot_vectors.dim() == 2
# foreground/background segmentation
mask_logits = self.inst_seg_net({'features': features, 'one_hot_vectors': one_hot_vectors})
# mask out Background points
foreground_coords, foreground_coords_mean, _ = F.logits_mask(
coords=features[:, :3, :], logits=mask_logits, num_points_per_object=self.num_points_per_object
)
# center regression
delta_coords = self.center_reg_net({'coords': foreground_coords, 'one_hot_vectors': one_hot_vectors})
foreground_coords = foreground_coords - delta_coords.unsqueeze(-1)
# box estimation
estimation = self.box_est_net({'coords': foreground_coords, 'one_hot_vectors': one_hot_vectors})
estimations = estimation.split([3, self.num_heading_angle_bins, self.num_heading_angle_bins,
self.num_size_templates, self.num_size_templates * 3], dim=-1)
# parse results
outputs = dict()
outputs['mask_logits'] = mask_logits
outputs['center_reg'] = foreground_coords_mean + delta_coords
outputs['center'] = estimations[0] + outputs['center_reg']
outputs['heading_scores'] = estimations[1]
outputs['heading_residuals_normalized'] = estimations[2]
outputs['heading_residuals'] = estimations[2] * (np.pi / self.num_heading_angle_bins)
outputs['size_scores'] = estimations[3]
size_residuals_normalized = estimations[4].view(-1, self.num_size_templates, 3)
outputs['size_residuals_normalized'] = size_residuals_normalized
outputs['size_residuals'] = size_residuals_normalized * self.size_templates
return outputs
def forward(self, inputs, cons=1, adaptation=False):
features = inputs['features']
one_hot_vectors = inputs['one_hot_vectors']
assert one_hot_vectors.dim() == 2
# foreground/background segmentation
mask_logits1, mask_logits2 = self.inst_seg_net(
{'features': features, 'one_hot_vectors': one_hot_vectors}, cons, adaptation)
mask_logits = (mask_logits1 + mask_logits2) / 2.0
# mask out Background points
foreground_coords, foreground_coords_mean, _ = F.logits_mask(
coords=features[:, :3, :], logits=mask_logits,
num_points_per_object=self.num_points_per_object
)
# center regression
delta_coords1, delta_coords2 = self.center_reg_net({'coords': foreground_coords,
'one_hot_vectors': one_hot_vectors}, cons, adaptation)
delta_coords = (delta_coords1 + delta_coords2) / 2.0
foreground_coords = foreground_coords - delta_coords.unsqueeze(-1)
# box estimation
estimation1, estimation2 = self.box_est_net({'coords': foreground_coords,
'one_hot_vectors': one_hot_vectors}, cons, adaptation)
estimation = (estimation1 + estimation2) / 2.0
estimations = estimation.split([3, self.num_heading_angle_bins,
self.num_heading_angle_bins,
self.num_size_templates,
self.num_size_templates * 3], dim=-1)
estimations1 = estimation1.split([3, self.num_heading_angle_bins,
self.num_heading_angle_bins,
self.num_size_templates,
self.num_size_templates * 3], dim=-1)
estimations2 = estimation2.split([3, self.num_heading_angle_bins,
self.num_heading_angle_bins,
self.num_size_templates,
self.num_size_templates * 3], dim=-1)
# parse results
outputs = dict()
outputs['mask_logits'] = mask_logits
outputs['mask_logits1'] = mask_logits1
outputs['mask_logits2'] = mask_logits2
outputs['center_reg'] = foreground_coords_mean + delta_coords
outputs['center_reg1'] = foreground_coords_mean + delta_coords1
outputs['center_reg2'] = foreground_coords_mean + delta_coords2
outputs['center'] = estimations[0] + outputs['center_reg']
outputs['center1'] = estimations1[0] + outputs['center_reg1']
outputs['center2'] = estimations2[0] + outputs['center_reg2']
outputs['heading_scores'] = estimations[1]
outputs['heading_residuals_normalized'] = estimations[2]
outputs['heading_residuals'] = estimations[2] * (np.pi / self.num_heading_angle_bins)
outputs['size_scores'] = estimations[3]
outputs['heading_scores1'] = estimations1[1]
outputs['heading_residuals_normalized1'] = estimations1[2]
outputs['heading_residuals1'] = estimations1[2] * (np.pi / self.num_heading_angle_bins)
outputs['size_scores1'] = estimations1[3]
outputs['heading_scores2'] = estimations2[1]
outputs['heading_residuals_normalized2'] = estimations2[2]
outputs['heading_residuals2'] = estimations2[2] * (np.pi / self.num_heading_angle_bins)
outputs['size_scores2'] = estimations2[3]
size_residuals_normalized = estimations[4].view(-1, self.num_size_templates, 3)
outputs['size_residuals_normalized'] = size_residuals_normalized
outputs['size_residuals'] = size_residuals_normalized * self.size_templates
size_residuals_normalized1 = estimations1[4].view(-1, self.num_size_templates, 3)
outputs['size_residuals_normalized1'] = size_residuals_normalized1
outputs['size_residuals1'] = size_residuals_normalized1 * self.size_templates
size_residuals_normalized2 = estimations2[4].view(-1, self.num_size_templates, 3)
outputs['size_residuals_normalized2'] = size_residuals_normalized2
outputs['size_residuals2'] = size_residuals_normalized2 * self.size_templates
return outputs