1 pose_generation.positional_sampling

pose_generation.positional_sampling

Direction and position sampling utilities for finite candidate generation.

This module samples candidate camera centers in the reference rig frame before they are transformed into world coordinates. It owns only the center prior: orientation, feasibility pruning, and candidate-table compaction are handled by neighboring pose_generation modules.

1.1 Theory

Raw directions are sampled either uniformly on the sphere, $u \sim \mathrm{Unif}(\mathbb{S}^2)$, or from a forward-biased PowerSpherical distribution, $p(u) \propto (1 + e_z^\top u)^\kappa$. Azimuth caps scale $\psi=\operatorname{atan2}(u_x,u_z)$ into the configured field $\Delta\psi$, while elevation caps map $u_y=\sin\theta$ into $[\sin\theta_{\min},\sin\theta_{\max}]$ without rejection. A radius $r\sim\mathcal{U}(r_{\min},r_{\max})$ then produces the reference-frame offset $o_r=r d_r$ and world center $c_w=T^w_r o_r$.

Position modes reinterpret the capped direction before the radius is applied. forward_local, local_refinement, and revisit_backtrack blend toward continuity priors; target_bearing_local blends toward the selected actor-visible target bearing; target_orbit makes balanced partial arcs at the current horizontal target standoff; lateral_target_bypass combines target bearing, signed lateral bypass, and bounded vertical offset. These target modes may use actor-visible target centers only, never GT target geometry.

1.2 Classes

Name	Description
PositionSampler	Sample candidate centers around a reference pose.

1.2.1 PositionSampler

pose_generation.positional_sampling.PositionSampler(config, *, device)

Sample candidate centers around a reference pose.

1.2.1.1 Methods

Name	Description
sample	Draw candidate centers and offsets in reference frame.

1.2.1.1.1 sample

pose_generation.positional_sampling.PositionSampler.sample(
    reference_pose,
    *,
    count,
    target_center_world,
)

Draw candidate centers and offsets in reference frame.

1.2.1.1.1.1 Parameters

Name	Type	Description	Default
reference_pose	`PoseTW`	`PoseTW` reference2world pose used as sampling origin.	required
count	int	Exact number of centers to attempt.	required
target_center_world	torch.Tensor \| None	Request-local actor-visible target, when required by the center family.	required

1.2.1.1.1.2 Returns

Name	Type	Description
	torch.Tensor	Tuple of `(centers_world, offsets_ref)` where both are `Tensor[N, 3]` with
	torch.Tensor	`N = count`. Offsets are in the reference frame before rotation into world.

# pose_generation.positional_sampling { #aria_nbv.pose_generation.positional_sampling } `pose_generation.positional_sampling` Direction and position sampling utilities for finite candidate generation. This module samples candidate camera centers in the reference rig frame before they are transformed into world coordinates. It owns only the center prior: orientation, feasibility pruning, and candidate-table compaction are handled by neighboring `pose_generation` modules. ## Theory {.doc-section .doc-section-theory} Raw directions are sampled either uniformly on the sphere, $u \sim \mathrm{Unif}(\mathbb{S}^2)$, or from a forward-biased PowerSpherical distribution, $p(u) \propto (1 + e_z^\top u)^\kappa$. Azimuth caps scale $\psi=\operatorname{atan2}(u_x,u_z)$ into the configured field $\Delta\psi$, while elevation caps map $u_y=\sin\theta$ into $[\sin\theta_{\min},\sin\theta_{\max}]$ without rejection. A radius $r\sim\mathcal{U}(r_{\min},r_{\max})$ then produces the reference-frame offset $o_r=r d_r$ and world center $c_w=T^w_r o_r$. Position modes reinterpret the capped direction before the radius is applied. `forward_local`, `local_refinement`, and `revisit_backtrack` blend toward continuity priors; `target_bearing_local` blends toward the selected actor-visible target bearing; `target_orbit` makes balanced partial arcs at the current horizontal target standoff; `lateral_target_bypass` combines target bearing, signed lateral bypass, and bounded vertical offset. These target modes may use actor-visible target centers only, never GT target geometry. ## Classes | Name | Description | | --- | --- | | [PositionSampler](#aria_nbv.pose_generation.positional_sampling.PositionSampler) | Sample candidate centers around a reference pose. | ### PositionSampler { #aria_nbv.pose_generation.positional_sampling.PositionSampler } ```python pose_generation.positional_sampling.PositionSampler(config, *, device) ``` Sample candidate centers around a reference pose. #### Methods | Name | Description | | --- | --- | | [sample](#aria_nbv.pose_generation.positional_sampling.PositionSampler.sample) | Draw candidate centers and offsets in reference frame. | ##### sample { #aria_nbv.pose_generation.positional_sampling.PositionSampler.sample } ```python pose_generation.positional_sampling.PositionSampler.sample( reference_pose, *, count, target_center_world, ) ``` Draw candidate centers and offsets in reference frame. ###### Parameters {.doc-section .doc-section-parameters} | Name | Type | Description | Default | |---------------------|---------------------------------------------------|-------------------------------------------------------------------------|------------| | reference_pose | [PoseTW](`efm3d.aria.pose.PoseTW`) | ``PoseTW`` reference2world pose used as sampling origin. | _required_ | | count | [int](`int`) | Exact number of centers to attempt. | _required_ | | target_center_world | [torch](`torch`).[Tensor](`torch.Tensor`) \| None | Request-local actor-visible target, when required by the center family. | _required_ | ###### Returns {.doc-section .doc-section-returns} | Name | Type | Description | |--------|-------------------------------------------|--------------------------------------------------------------------------------| | | [torch](`torch`).[Tensor](`torch.Tensor`) | Tuple of ``(centers_world, offsets_ref)`` where both are ``Tensor[N, 3]`` with | | | [torch](`torch`).[Tensor](`torch.Tensor`) | ``N = count``. Offsets are in the reference frame before rotation into world. |