Hestia

1 Hestia: Hierarchical Continuous NBV Bridge

Primary source. Hestia: Voxel-Face-Aware Hierarchical Next-Best-View Acquisition for Efficient 3D Reconstruction [1].

Local source. main.tex and section sources under tex-src/arXiv-Hestia/sec/.

Related ARIA-NBV pages. GenNBV, RL planning, and roadmap.

1.1 Core contribution

Hestia is the most relevant continuous-policy bridge in the current corpus. It remains coverage-driven, but it improves the continuous NBV control structure by decomposing action selection into a look-at proposal and a feasible camera-position policy, and by tracking directional voxel-face visibility [1].

For ARIA-NBV, the transferable idea is not the coverage reward. It is the factorization:

\[ \text{target/look-at proposal} \;\longrightarrow\; \text{feasible camera pose}. \]

1.2 Verified paper signals

signal	source-backed detail	ARIA-NBV relevance
Face-aware state	Hestia tracks directional visibility of voxel faces rather than only occupancy.	Inspires directional observation histograms for semi-dense points, EVL voxels, or target crops.
Hierarchical action	The controller predicts a look-at point and a 3-DoF position; yaw/pitch are derived from the look-at relation.	Future bridge from finite candidates to target-then-pose continuous control.
Feasibility projection	Predicted positions are adjusted to collision-free locations.	Mirrors ARIA-NBV’s need for hard validity masks and feasible candidate transitions.
Close-greedy reward	The paper uses immediate face-coverage gains with a small discount factor.	Supports short-horizon evidence gates before long-horizon RL claims.
Intermediate supervision	The look-at proposal is supervised by a target derived from uncaptured surface geometry.	Suggests supervised target/missing-surface proposal heads after target-RRI labels are trusted.

Directional face visibility can be summarized as:

\[ F_t = f_t \lor F_{t-1}, \qquad f_t(v_i,j) = \mathbb{1}\!\left( \frac{a'_t-p_{v_i}}{\lVert a'_t-p_{v_i}\rVert} \cdot n_{i,j} > 0 \right). \]

1.3 ARIA-NBV adoption

Stretch/bridge: use Hestia to design a later continuous target-then-pose actor, not as the first quantitative thesis baseline.
Proposal/diagnostic: add directional observation features or incidence histograms to candidate reports.
Validity: keep feasibility projection and invalid-action handling as hard constraints rather than reward-only penalties.

1.4 Do not adopt

Do not replace RRI or target RRI with face coverage.
Do not make continuous actor-critic training mandatory for the thesis result.
Do not transfer object-centric Objaverse coverage results as evidence for ASE/EFM target-conditioned NBV.

1.5 Open risks / caveats

Directional face coverage is still a proxy and may disagree with Chamfer-style quality.
Continuous policies need simulator fidelity, collision handling, and reward throughput that ARIA-NBV does not yet have.
Hestia is most useful after finite-candidate oracle rollouts and Q_H learning establish the quality-driven baseline.

References

[1]

C.-Y. Lu et al., “Hestia: Voxel-face-aware hierarchical next-best-view acquisition for efficient 3D reconstruction,” in Proceedings of the IEEE/CVF winter conference on applications of computer vision, 2026. Available: https://openaccess.thecvf.com/content/WACV2026/papers/Lu_Hestia_Voxel-Face-Aware_Hierarchical_Next-Best-View_Acquisition_for_Efficient_3D_Reconstruction_WACV_2026_paper.pdf

--- title: "Hestia" phase: thesis audience: public status: current owner: jan format: html --- ## Hestia: Hierarchical Continuous NBV Bridge {#hestia} **Primary source.** Hestia: Voxel-Face-Aware Hierarchical Next-Best-View Acquisition for Efficient 3D Reconstruction [@Hestia-lu2026]. **Local source.** [`main.tex`](../../literature/tex-src/arXiv-Hestia/main.tex) and section sources under [`tex-src/arXiv-Hestia/sec/`](../../literature/tex-src/arXiv-Hestia/sec/). **Related ARIA-NBV pages.** [GenNBV](gen_nbv.qmd), [RL planning](rl_planning.qmd), and [roadmap](../thesis/roadmap.qmd). ### Core contribution Hestia is the most relevant continuous-policy bridge in the current corpus. It remains coverage-driven, but it improves the continuous {{< gls next-best-view >}} control structure by decomposing action selection into a look-at proposal and a feasible camera-position policy, and by tracking directional voxel-face visibility [@Hestia-lu2026]. For ARIA-NBV, the transferable idea is not the coverage reward. It is the factorization: $$ \text{target/look-at proposal} \;\longrightarrow\; \text{feasible camera pose}. $$ ### Verified paper signals | signal | source-backed detail | ARIA-NBV relevance | |---|---|---| | Face-aware state | Hestia tracks directional visibility of voxel faces rather than only occupancy. | Inspires directional observation histograms for semi-dense points, EVL voxels, or target crops. | | Hierarchical action | The controller predicts a look-at point and a 3-DoF position; yaw/pitch are derived from the look-at relation. | Future bridge from finite candidates to target-then-pose continuous control. | | Feasibility projection | Predicted positions are adjusted to collision-free locations. | Mirrors ARIA-NBV's need for hard validity masks and feasible candidate transitions. | | Close-greedy reward | The paper uses immediate face-coverage gains with a small discount factor. | Supports short-horizon evidence gates before long-horizon RL claims. | | Intermediate supervision | The look-at proposal is supervised by a target derived from uncaptured surface geometry. | Suggests supervised target/missing-surface proposal heads after target-RRI labels are trusted. | Directional face visibility can be summarized as: $$ F_t = f_t \lor F_{t-1}, \qquad f_t(v_i,j) = \mathbb{1}\!\left( \frac{a'_t-p_{v_i}}{\lVert a'_t-p_{v_i}\rVert} \cdot n_{i,j} > 0 \right). $$ ### ARIA-NBV adoption - **Stretch/bridge:** use Hestia to design a later continuous target-then-pose actor, not as the first quantitative thesis baseline. - **Proposal/diagnostic:** add directional observation features or incidence histograms to candidate reports. - **Validity:** keep feasibility projection and invalid-action handling as hard constraints rather than reward-only penalties. ### Do not adopt - Do not replace {{< gls relative-reconstruction-improvement >}} or {{< gls target-specific-rri >}} with face coverage. - Do not make continuous actor-critic training mandatory for the thesis result. - Do not transfer object-centric Objaverse coverage results as evidence for ASE/EFM target-conditioned NBV. ### Open risks / caveats - Directional face coverage is still a proxy and may disagree with Chamfer-style quality. - Continuous policies need simulator fidelity, collision handling, and reward throughput that ARIA-NBV does not yet have. - Hestia is most useful after finite-candidate oracle rollouts and {{< gls finite-horizon-q-function >}} learning establish the quality-driven baseline.