NBV Background

1 Classical Methods: Geometric Heuristics

Hand-crafted criteria include:

• Volumetric uncertainty (unseen voxels)
• Information gain (entropy reduction)
• Surface coverage [1]

2 Coverage-Based Learning: GenNBV

GenNBV (literature-review) [2] introduced learning-based NBV with:

• RL-based training with continuous 5DoF actions
• Coverage-focused rewards for geometric completeness

Limitation: Still optimizes geometric coverage, not reconstruction quality. Primitive non-pre-trained architecture.

3 Direct Quality Optimization: VIN-NBV

VIN-NBV (literature-review) [1] directly optimizes Relative Reconstruction Improvement (RRI):

\[\text{RRI}(v) = \frac{\text{Quality}(R_{t+1}) - \text{Quality}(R_t)}{\text{Quality}(R_{\text{complete}})}\]

• Predicts actual reconstruction quality improvement
• Trained via imitation learning on oracle RRI values using CORAL loss
• Uses surface metrics (Chamfer distance, F-score) as quality measures
• Demonstrates ≈30% improvement over coverage-based methods

Limitation: Custom CNN backbone limits generalization to complex scenes, doesn’t leverage ego-centric priors / modalities. Greedy selection of randomly sampled views based on predicted RRI.

References

[1]

N. Frahm et al., “VIN-NBV: A view introspection network for next-best-view selection.” 2025. Available: https://arxiv.org/abs/2505.06219

[2]

X. Chen, Q. Li, T. Wang, T. Xue, and J. Pang, “GenNBV: Generalizable next-best-view policy for active 3D reconstruction.” 2024. Available: https://arxiv.org/abs/2402.16174