PB-NBV

1 PB-NBV: Projection-Based Candidate Proposals

Primary source. PB-NBV: Efficient Projection-Based Next-Best-View Planning Framework for Reconstruction of Unknown Objects [1].

Local source. jzz_2025_ral_resub.tex and sections/method.tex.

Related ARIA-NBV pages. NBV background, RRI theory, oracle RRI API, and Hestia.

1.1 Core contribution

PB-NBV is an efficient object-centric NBV framework that avoids dense per-candidate ray casting by fitting compact ellipsoid proxies to classified voxel clusters and scoring candidates through 2D projection [1]. It is not a replacement for RRI. It is a proposal heuristic for reducing candidate-evaluation cost.

1.2 Verified paper signals

signal	source-backed detail	ARIA-NBV relevance
Candidate region	The method samples candidate poses on a partial hemisphere around the object.	Useful as a reminder that candidate geometry is a controllable proposal distribution.
Voxel taxonomy	The method separates empty/free, occupied, unknown, frontier, and unclassified voxels.	Maps to semi-dense surface support, EVL free-space evidence, and missing/frontier proxies.
Ellipsoid proxy	Occupied and frontier voxel clusters are approximated by minimum-volume enclosing ellipsoids.	Suggests cheap proxy geometry before expensive oracle rendering/mesh-distance scoring.
Projection score	Candidate views are scored by projected frontier/occupied evidence, with simple depth-rank weighting.	Useful as a shortlist score or diagnostic, not as the final thesis utility.
Continuity	A global partitioning strategy supports scan continuity and registration stability.	Motivates motion/path and overlap constraints in multi-step ARIA-NBV rollouts.

PB-NBV’s voxel-state distinction can be translated into ARIA-NBV as:

PB-NBV state	ARIA-NBV analogue
empty/free	EVL free-space evidence or known-free ray support
occupied	semi-dense points or EVL occupancy evidence
unknown	unobserved/incomplete regions inside candidate support
frontier	boundary between observed support and unknown target/scene regions
unclassified	out-of-support or insufficient-evidence regions

1.3 ARIA-NBV adoption

Proposal/diagnostic: build cheap frontier/free-space projection scores to shortlist candidates before oracle RRI evaluation.
Candidate strategy: compare LookAway, LookAt target, frontier, random, and PB-NBV-style projection strategies by valid ratio, target visibility, and RRI histograms.
Rollout constraint: preserve motion continuity, overlap, and mask diagnostics instead of optimizing utility alone.

The transfer pattern is:

Code

flowchart LR
  A["Semi-dense / EVL evidence"] --> B["Frontier or missing-support proxy"]
  B --> C["Projection score"]
  C --> D["Candidate shortlist"]
  D --> E["Oracle scene + target RRI"]

flowchart LR
  A["Semi-dense / EVL evidence"] --> B["Frontier or missing-support proxy"]
  B --> C["Projection score"]
  C --> D["Candidate shortlist"]
  D --> E["Oracle scene + target RRI"]

1.4 Do not adopt

Do not replace scene or target RRI with projected frontier area.
Do not assume turntable/object-centric geometry for egocentric indoor ASE snippets.
Do not treat ellipsoid depth rank as ground-truth visibility.
Do not assume clean structured-light depth or ICP-style registration inputs.

1.5 Open risks / caveats

Projection scores can favor large visible areas rather than reconstruction-quality improvement.
Frontier proxies need calibration against oracle RRI and target RRI.
PB-NBV is best read as a fast proposal mechanism, not as the thesis objective.

References

[1]

Z. Jia, Y. Li, Q. Hao, and S. Zhang, “PB-NBV: Efficient projection-based next-best-view planning framework for reconstruction of unknown objects,” IEEE Robotics and Automation Letters, vol. 10, no. 7, pp. 7444–7451, 2025, doi: 10.1109/LRA.2025.3573631.

--- title: "PB-NBV" phase: thesis audience: public status: current owner: jan format: html --- ## PB-NBV: Projection-Based Candidate Proposals {#pb-nbv-literature} **Primary source.** [PB-NBV: Efficient Projection-Based Next-Best-View Planning Framework for Reconstruction of Unknown Objects](https://doi.org/10.1109/LRA.2025.3573631) [@PB-NBV-jia2025]. **Local source.** [`jzz_2025_ral_resub.tex`](../../literature/tex-src/arXiv-PB-NBV/jzz_2025_ral_resub.tex) and [`sections/method.tex`](../../literature/tex-src/arXiv-PB-NBV/sections/method.tex). **Related ARIA-NBV pages.** [NBV background](../theory/nbv_background.qmd), [RRI theory](../theory/rri_theory.qmd), [oracle RRI API](../../reference/rri_metrics.oracle_rri.qmd), and [Hestia](hestia.qmd). ### Core contribution PB-NBV is an efficient object-centric {{< gls next-best-view >}} framework that avoids dense per-candidate ray casting by fitting compact ellipsoid proxies to classified voxel clusters and scoring candidates through 2D projection [@PB-NBV-jia2025]. It is not a replacement for {{< gls relative-reconstruction-improvement >}}. It is a proposal heuristic for reducing candidate-evaluation cost. ### Verified paper signals | signal | source-backed detail | ARIA-NBV relevance | |---|---|---| | Candidate region | The method samples candidate poses on a partial hemisphere around the object. | Useful as a reminder that candidate geometry is a controllable proposal distribution. | | Voxel taxonomy | The method separates empty/free, occupied, unknown, frontier, and unclassified voxels. | Maps to semi-dense surface support, EVL free-space evidence, and missing/frontier proxies. | | Ellipsoid proxy | Occupied and frontier voxel clusters are approximated by minimum-volume enclosing ellipsoids. | Suggests cheap proxy geometry before expensive oracle rendering/mesh-distance scoring. | | Projection score | Candidate views are scored by projected frontier/occupied evidence, with simple depth-rank weighting. | Useful as a shortlist score or diagnostic, not as the final thesis utility. | | Continuity | A global partitioning strategy supports scan continuity and registration stability. | Motivates motion/path and overlap constraints in multi-step ARIA-NBV rollouts. | PB-NBV's voxel-state distinction can be translated into ARIA-NBV as: | PB-NBV state | ARIA-NBV analogue | |---|---| | empty/free | EVL free-space evidence or known-free ray support | | occupied | semi-dense points or EVL occupancy evidence | | unknown | unobserved/incomplete regions inside candidate support | | frontier | boundary between observed support and unknown target/scene regions | | unclassified | out-of-support or insufficient-evidence regions | ### ARIA-NBV adoption - **Proposal/diagnostic:** build cheap frontier/free-space projection scores to shortlist candidates before oracle RRI evaluation. - **Candidate strategy:** compare LookAway, LookAt target, frontier, random, and PB-NBV-style projection strategies by valid ratio, target visibility, and RRI histograms. - **Rollout constraint:** preserve motion continuity, overlap, and {{< gls validity-mask >}} diagnostics instead of optimizing utility alone. The transfer pattern is: ```{mermaid} flowchart LR A["Semi-dense / EVL evidence"] --> B["Frontier or missing-support proxy"] B --> C["Projection score"] C --> D["Candidate shortlist"] D --> E["Oracle scene + target RRI"] ``` ### Do not adopt - Do not replace scene or target RRI with projected frontier area. - Do not assume turntable/object-centric geometry for egocentric indoor ASE snippets. - Do not treat ellipsoid depth rank as ground-truth visibility. - Do not assume clean structured-light depth or ICP-style registration inputs. ### Open risks / caveats - Projection scores can favor large visible areas rather than reconstruction-quality improvement. - Frontier proxies need calibration against {{< gls oracle-rri >}} and {{< gls target-specific-rri >}}. - PB-NBV is best read as a fast proposal mechanism, not as the thesis objective.