LiTe-GS: Oracle-Efficient Next Best View Selection for 3D Gaussian Splatting

1. Introduction

• Volumetric rendering supports photorealistic novel-view synthesis, but remains computationally demanding for latency-sensitive robotics and real-time perception.Examples include autonomous navigation and simultaneous localization and mapping.Source: §1 · p. 1
• 3DGS offers reconstruction quality comparable to or exceeding NeRF and Plenoxels while reducing training time and enabling real-time rendering.These properties motivate its use in robotics tasks including mapping, navigation, and manipulation.Source: §1 · p. 1
• View selection must identify informative frames that reduce uncertainty in 3DGS parameters without the cost and redundancy of processing every available view.Information-driven methods face a trade-off between candidate-oracle evaluations and selection quality.Source: §1 · p. 2
• LiTe-GS is an oracle-efficient next-best-view framework that balances efficiency and information gain with provable approximation guarantees.It reduces Fisher-oracle evaluations through randomized subset selection, with the efficiency–approximation trade-off controlled by ε.Source: §1 · p. 2

2.1. 3D Gaussian Splatting

• 3DGS represents scenes with anisotropic Gaussians optimized by differentiable rasterization, while related work integrates Gaussians into SLAM and mapping systems.Photo-SLAM [13] combines Gaussians with ORB-SLAM; GS-ICP SLAM [9] and RTG-SLAM [28] use ICP-based registration, and other frameworks incorporate LiDAR [12] [18] [31].Source: §2.1 · pp. 2–3
• Prior work leaves efficient view selection for Gaussian reconstruction comparatively underexplored, as many methods assume fixed training views and focus on downstream tasks.The paper positions uncertainty-aware selection for 3DGS as the gap it addresses.Source: §2.1 · p. 3

2.2. Active View Selection and Next-Best-View Planning

• FisherRF [14] selects views by expected information gain, and POP-GS [34] uses a block-diagonal Fisher approximation, but repeated candidate evaluations remain necessary.LiTe-GS retains the information-driven objective while using randomized sampling to reduce oracle evaluations and expose an efficiency–approximation trade-off.Source: §2.2 · p. 3

3.1. 3D Gaussian Splatting

• 3DGS models a scene as Gaussian primitives with position, rotation, anisotropic scale, opacity, and view-dependent appearance, each defining a spatial density.The covariance controls each Gaussian’s spatial extent and orientation.Source: §3.1 · p. 3
• Training views determine the 3DGS parameters through gradient-based optimization of a weighted L1 and SSIM loss between rendered and ground-truth images.The representation uses differentiable rasterization and alpha blending of ordered Gaussian contributions along each pixel’s viewing ray.Source: §3.1 · p. 3

3.2. Uncertainty Reduction in Active View Selection

• Active acquisition begins with seed views and a prior parameter estimate; a local linearization around that estimate supports approximating posterior covariance.The rendered observation for a candidate view is modeled using the differentiable renderer and isotropic measurement noise.Source: §3.2 · p. 4
• The method uses selected-view Fisher information with a block-diagonal approximation across Gaussian primitives to make storage and incremental updates tractable.The approximation preserves intra-splat information; the full information matrix would be prohibitive for scenes with millions of parameters.Source: §3.2 · p. 4

3.3. View Selection as Binary Optimization

• Given candidate viewpoints, the objective is to choose a subset that maximally reduces uncertainty in the estimated 3DGS parameters.Each candidate contributes additive information, evaluated against a scalar uncertainty metric over the accumulated information matrix.Source: §3.3 · p. 4
• Binary variables encode candidate selection under a fixed view-cardinality constraint, yielding a combinatorial optimization problem that is generally intractable to solve exactly.This motivates efficient approximation algorithms.Source: §3.3 · p. 4

4. Methodology

• LiTe-GS retains POP-GS’s Fisher-information objective, randomizes candidate sampling, and leaves subsequent 3DGS optimization unchanged.The efficiency analysis concerns the view-selection module rather than end-to-end 3DGS training time.Source: §4 · pp. 4–5

4.1. Uncertainty Metric

• Posterior covariance characterizes Gaussian-parameter uncertainty, and differential entropy increases monotonically with the covariance log-determinant.The method defines ρ(v | Sq) as the marginal gain from adding candidate view v to selected set Sq; computing it accumulates that view’s Fisher-information contribution and is one oracle evaluation.Source: §4.1 · p. 5

4.2. Performance Analysis

• The paper analyzes Algorithm 1 for oracle complexity and theoretical optimality guarantees.The approximation analysis compares its solution with the optimal solution under the stated submodularity conditions.Source: §4.2 · p. 5, §4.2.2 · p. 5

4.2.1. Oracle Efficiency Analysis

• Classical greedy selection requires O(MK) oracle evaluations to choose K views from M candidates, evaluating all remaining candidates at each iteration.LiTe-GS uses randomized sampling and lazy marginal evaluation, and its analyzed oracle complexity is independent of K.Source: §4.2.1 · p. 5

4.2.2. Approximation Guarantees

• Under a normalized, non-negative, monotone submodular objective, the method achieves a near-optimal solution in expectation with oracle complexity independent of K and dependent on ε.Increasing ε reduces oracle evaluations at the expense of a weaker approximation guarantee; POP-GS uses exhaustive oracle evaluations and lacks this trade-off.Source: §4.2.2 · p. 5

5. Experiments

• The experiments evaluate LiTe-GS through qualitative and quantitative results.The section describes datasets, baseline comparisons, experimental setup, and evaluation metrics.Source: §5 · p. 6
• Experiments use ε = 0.1 in Algorithm 1 unless otherwise stated.Source: §5 · p. 6
• The experimental framework extends the FisherRF repository to incorporate LiTe-GS and the block-diagonal Fisher Information approximation.Source: §5 · p. 6

5.1. Experimental Setup and Dataset

• Blender supplies eight synthetic scenes with 100 training and 200 test images each, while Mip-NeRF 360 supplies complex, high-resolution real-world scenes.The comparison uses POP-GS and random selection; FisherRF is not a separate baseline because its diagonal-Fisher formulation is a special case of POP-GS.Source: §5.1 · p. 6
• The evaluation metrics are PSNR, SSIM, and LPIPS.Source: §5.1 · p. 6

5.1.1. Setup 1

• Setup 1 starts with two images and adds three per acquisition until 11 training images, training for 100v iterations after each acquisition.Source: §5.1.1 · p. 6

5.1.2. Setup 2

• Setup 2, conducted only on Blender, starts with five images and adds four per acquisition to reach 21, using 300v iterations between acquisitions.The setup totals 21,000 training iterations.Source: §5.1.2 · p. 6

5.1.3. Setup 3

• Setup 3, conducted only on Mip-NeRF 360, starts with three images and adds four per acquisition to reach 15, training for 100v iterations after each acquisition.The setup totals 11,000 training iterations.Source: §5.1.3 · p. 6

5.2.1. Rendered Image Quality

• LiTe-GS achieves reconstruction quality comparable to POP-GS across most scenes and substantially reduces Fisher-oracle evaluations across experimental settings.The evaluation reports average performance over three scenes per dataset; Table 3 compares oracle evaluations with POP-GS.Source: §5.2.1 · p. 6, §5.2.2 · p. 7, Table 3 · p. 7
• Randomized candidate sampling occasionally lets LiTe-GS outperform POP-GS, but these cases are not claimed as a systematic quality advantage.The authors attribute such cases to exploring different candidate subsets rather than always making locally greedy choices.Source: §5.2.1 · pp. 6–7

5.2.2. Oracle Efficiency Results

• Oracle efficiency is measured by Fisher-information evaluations, and LiTe-GS restricts evaluations to a sampled subset rather than scoring every candidate at each iteration.For conservative accounting, all sampled candidates count as evaluations, without credit for additional savings from lazy evaluation.Source: §5.2.2 · p. 7
• With ε = 0.1, increasing K from 3 to 6 raises POP-GS evaluations from 275 to 549 while LiTe-GS stays nearly constant at 210, 209, and 208.The authors interpret this Setup 4 result as demonstrating the predicted K-independence of LiTe-GS oracle complexity.Source: §5.2.2 · p. 7, Table 4 · p. 8
• At K = 6, increasing ε reduces LiTe-GS evaluations from 208 to 108 and 44, yielding 80.42% and 91.99% reductions with an approximation-quality trade-off.The ε-controlled trade-off lets users tune oracle expenditure to the desired accuracy level.Source: §5.2.2 · p. 7

5.3. Ablation Study

• LiTe-GS resamples candidate subsets at each acquisition step, whereas SSG samples once per stage and keeps the subset fixed; both use the same subset size for each ε.This comparison isolates dynamic versus static candidate restriction.Source: §5.3 · p. 8
• On Blender (Ship), a fixed subset can occasionally outperform LiTe-GS, while on Mip-NeRF 360 (Garden), LiTe-GS consistently outperforms SSG across evaluated ε values.Together, the studies do not guarantee better reconstruction for every realization, but indicate more robust candidate exploration than permanent restriction to one subset.Source: §5.3 · p. 8, Table 6 · p. 8
• Increasing ε reduces candidate evaluations while permitting a controlled approximation-quality trade-off, giving LiTe-GS a way to tune oracle expenditure to a desired accuracy level.This contrasts with exhaustive POP-GS selection, which does not provide that tuning mechanism.Source: §5.3 · p. 8

• LiTe-GS formulates next-best-view selection as cardinality-constrained maximization of Fisher-information gain for efficient approximate view selection.Source: §6 · p. 8
• A randomized strategy restricts computation to sampled candidate subsets at each iteration, reducing oracle evaluations with provable approximation guarantees.Source: §6 · p. 8
• Experiments show reconstruction quality comparable to state-of-the-art alongside significantly fewer Fisher-information oracle evaluations.Source: §6 · p. 8