Decodes features into delta values using convolutions.
| 13 | |
| 14 | |
| 15 | class DirectPredictionHead(nn.Module): |
| 16 | """Decodes features into delta values using convolutions.""" |
| 17 | |
| 18 | def __init__(self, feature_dim: int, num_layers: int) -> None: |
| 19 | """Initialize DirectGaussianPredictor. |
| 20 | |
| 21 | Args: |
| 22 | feature_dim: Number of input features. |
| 23 | num_layers: The number of layers of Gaussians to predict. |
| 24 | """ |
| 25 | super().__init__() |
| 26 | self.num_layers = num_layers |
| 27 | |
| 28 | # 14 is 3 means, 3 scales, 4 quaternions, 3 colors and 1 opacity |
| 29 | self.geometry_prediction_head = nn.Conv2d(feature_dim, 3 * num_layers, 1) |
| 30 | self.geometry_prediction_head.weight.data.zero_() |
| 31 | assert self.geometry_prediction_head.bias is not None |
| 32 | self.geometry_prediction_head.bias.data.zero_() |
| 33 | |
| 34 | self.texture_prediction_head = nn.Conv2d(feature_dim, (14 - 3) * num_layers, 1) |
| 35 | self.texture_prediction_head.weight.data.zero_() |
| 36 | assert self.texture_prediction_head.bias is not None |
| 37 | self.texture_prediction_head.bias.data.zero_() |
| 38 | |
| 39 | def forward(self, image_features: ImageFeatures) -> torch.Tensor: |
| 40 | """Predict deltas for 3D Gaussians. |
| 41 | |
| 42 | Args: |
| 43 | image_features: Image features from decoder. |
| 44 | |
| 45 | Returns: |
| 46 | The predicted deltas for Gaussian attributes. |
| 47 | """ |
| 48 | delta_values_geometry = self.geometry_prediction_head(image_features.geometry_features) |
| 49 | delta_values_texture = self.texture_prediction_head(image_features.texture_features) |
| 50 | delta_values_geometry = delta_values_geometry.unflatten(1, (3, self.num_layers)) |
| 51 | delta_values_texture = delta_values_texture.unflatten(1, (14 - 3, self.num_layers)) |
| 52 | delta_values = torch.cat([delta_values_geometry, delta_values_texture], dim=1) |
| 53 | return delta_values |