ObjView-Bench: Rethinking Difficulty and Deployment for Object-Centric View Planning

Sicong Pan1, Hao Hu2, Xuying Huang1, Benno Wingender1, Maren Bennewitz1
1University of Bonn, 2Independent Researcher

A difficulty-aware and deployment-oriented benchmark for object-centric view planning in active geometric 3D reconstruction.

ObjView-Bench tabletop robot deployment teaser.

ObjView-Bench studies how object geometry, view budget, and reachable-view constraints change view-planning behavior.

Abstract

Object-centric view planning is a core component of active geometric 3D reconstruction in robotics, yet existing evaluations often conflate object complexity, planning difficulty, budget assumptions, and physical reachability constraints. As a result, conclusions drawn from idealized view-planning evaluations may not reliably predict performance under realistic reconstruction settings.

We introduce ObjView-Bench, an evaluation framework for rethinking difficulty and deployment in object-centric view planning. First, we disentangle three quantities underlying view-planning evaluation: omnidirectional self-occlusion as an object-side attribute, observation saturation difficulty, and protocol-dependent planning difficulty defined through a set-cover formulation.

This separation supports controlled dataset construction, analysis of slow-saturation objects, and a case study showing that planning difficulty-aware sampling can improve learned view planners. We further design deployment-oriented evaluation protocols that reveal how budget regimes and reachable-view constraints alter method behavior. Across classical, learned, and hybrid planners, ObjView-Bench shows that difficulty, budget, and reachability constraints substantially change method rankings and failure modes.

Benchmark Overview

  • Difficulty semantics. ObjView-Bench separates omnidirectional self-occlusion, observation saturation difficulty, and protocol-dependent planning difficulty.
  • Difficulty-characterized object pool. The benchmark contains 5,691 manually reviewed objects, including 5,523 main-pool objects and 168 slow-saturation diagnostic objects.
  • Deployment-oriented evaluation. Methods are evaluated with online RGB-D observations, normalized surface coverage, path cost, runtime, multiple view-budget protocols, and reachable-view constraints.
  • Representative planner coverage. The study compares classical, learned, and hybrid planners under shared difficulty and deployment semantics.

Observation Saturation Difficulty

Objects saturate at different rates: simple geometries reveal their observable surface from a few generic views, while concavities, cavities, and inter-part occlusions require denser and more specific viewpoints.

Observation saturation behavior.

Observation saturation. Marginal gain curves motivate the 128-view planning protocol.

Planning difficulty distribution.

Planning difficulty. The main pool has a long-tailed planning-difficulty distribution.

Saturation difficulty versus planning difficulty.

Difficulty disentanglement. Saturation and planning difficulty are related but not equivalent.

Key Findings

ObjView-Bench is designed to study how different planning paradigms behave under difficulty, budget, and reachability shifts rather than to produce a single fixed-protocol leaderboard. The experiments show that method rankings are strongly budget-dependent: online view-selection methods perform well under small budgets, while coverage-oriented planning becomes stronger with larger budgets.

Reachability constraints substantially reduce absolute surface coverage and interact with each method's action representation. Classical and hybrid planners remain robust after normalization by reachable observable surface, whereas learned planners can be sensitive when their predicted actions become infeasible.

Slow-saturation objects reveal a gap between map stabilization and long-horizon observable-surface saturation. Planning difficulty also acts as a useful learning signal: difficulty-aware sampling improves learned view-set prediction on medium- and high-difficulty objects without changing the model architecture.

Planning difficulty-aware split.

Balanced sampling. Difficulty-aware sampling increases exposure to medium- and high-difficulty objects.

Raw versus balanced training effect.

Training effect. Balanced sampling improves achieved voxel coverage across the planning-difficulty axis.

Lessons for Future View-Planning Methods

  1. Scaling learned view-set prediction: larger candidate view sets and longer horizons expose a harder regime for learned set-prediction methods.
  2. Reachability-aware learning: constrained view spaces interact strongly with learned action representations and should be incorporated into training and scoring.
  3. Hybrid geometry learning and explicit planning: completion-guided planning is robust under budget and reachability shifts, but the explicit planning stage should be accelerated.
  4. Slow-saturation and budget allocation: hard-to-reveal surfaces create persistent marginal gains even after the observed map appears stable.
  5. Planning difficulty as a learning signal: difficulty can guide dataset construction, curriculum learning, and active learning.

BibTeX

@article{pan2026objviewbench,
  title         = {ObjView-Bench: Rethinking Difficulty and Deployment for Object-Centric View Planning},
  author        = {Pan, Sicong and Hu, Hao and Huang, Xuying and Wingender, Benno and Bennewitz, Maren},
  journal       = {arXiv preprint arXiv:2605.10707},
  year          = {2026}
}