controller

controller#

class pointrix.controller.gs.DensificationController(cfg: dict | DictConfig | None = None, *args, **kwargs)#

Bases: BaseDensificationController

The controller class for densifying the point cloud.

class Config(prune_interval: int = 100, min_opacity: float = 0.005, duplicate_interval: int = 100, densify_start_iter: int = 500, densify_stop_iter: int = 15000, densify_grad_threshold: float = 0.0002, optimizer_name: str = 'optimizer_1', split_num: int = 2, control_module: str = 'point_cloud', percent_dense: float = 0.01, opacity_reset_interval: int = 3000, normalize_grad: bool = True)#

Bases: Config

Parameters:

  • split_num (int) – The number of points to split.

  • control_module (str) – The control module.

  • percent_dense (float) – The percentage of dense points.

  • opacity_reset_interval (int) – The interval to reset the opacity.

  • densify_grad_threshold (float) – The controller gradient threshold.

  • min_opacity (float) – The minimum opacity.

  • normalize_grad (bool) – Whether to normalize the gradient.

accumulate_viewspace_grad(uv_points: Tensor) → Tensor#

Accumulate viewspace gradients for batch.

Parameters:

uv_points (torch.Tensor) – The view space points.

Returns:

The viewspace gradients.

Return type:

torch.Tensor

densify_clone(grads: Tensor) → None#

Densify the point cloud by cloning.

Parameters:

grads (torch.Tensor) – The gradients.

densify_split(grads: Tensor) → None#

Densify the point cloud by splitting.

Parameters:

grads (torch.Tensor) – The gradients.

generate_clone_mask(grads: Tensor) → Tensor#

Generate the mask for cloning. :param grads: The gradients. :type grads: torch.Tensor

Returns:

The mask for cloning.

Return type:

torch.Tensor

generate_split_mask(grads: Tensor) → Tensor#

Generate the mask for splitting.

Parameters:

grads (torch.Tensor) – The gradients.

new_pos_scale(mask: Tensor) → Tuple[Tensor, Tensor]#

Generate new position and scaling for splitting.

Parameters:

mask (torch.Tensor) – The mask for splitting.

Returns:

The new position and scaling.

Return type:

tuple[torch.Tensor, torch.Tensor]

prune(**kwargs) → None#

Prune the point cloud.

Parameters:

step (int) – The current step.

prune_postprocess(valid_points_mask)#

Postprocess after pruning.

Parameters:

valid_points_mask (torch.Tensor) – The mask for valid points.

reset_controller_state() → None#

Reset the controller state.

reset_opacity(reset_scele=0.01) → None#

Reset the opacity of the point cloud.

Parameters:

reset_scele (float) – The reset scale.

Contents