cs.CVApr 6, 2026

Coverage Optimization for Camera View Selection

Authors: Timothy Chen, Adam Dai, Maximilian Adang, Grace Gao, Mac Schwager

Organizations: Stanford University

Abstract

What makes a good viewpoint? The quality of the data used to learn 3D reconstructions is crucial for enabling efficient and accurate scene modeling. We study the active view selection problem and develop a principled analysis that yields a simple and interpretable criterion for selecting informative camera poses. Our key insight is that informative views can be obtained by minimizing a tractable approximation of the Fisher Information Gain, which reduces to favoring viewpoints that cover geometry that has been insufficiently observed by past cameras. This leads to a lightweight coverage-based view selection metric that avoids expensive transmittance estimation and is robust to noise and training dynamics. We call this metric COVER (Camera Optimization for View Exploration and Reconstruction). We integrate our method into the Nerfstudio framework and evaluate it on real datasets within fixed and embodied data acquisition scenarios. Across multiple datasets and radiance-field baselines, our method consistently improves reconstruction quality compared to state-of-the-art active view selection methods. Additional visualizations and our Nerfstudio package can be found at https://chengine.github.io/nbv_gym/.

Figures & tables

Figure 1: COVER is a simple, performant, and from-first-principles view selection metric that can be batch-queried in real-time and rendered into an image for visualization. The view metric measures the difference between perspective candidate cameras and cameras in the training dataset. Candidate viewpoints that cover large parts of the scene and are under-covered by the training cameras are selected and added to the training dataset to improve the quality of the \(3\) D scene reconstruction. — Figure 1: COVER is a simple, performant, and from-first-principles view selection metric that can be batch-queried in real-time and rendered into an image for visualization. The view metric measures the difference between perspective candidate cameras and cameras in the training dataset. Candidate viewpoints that cover large parts of the scene and are under-covered by the training cameras are selected and added to the training dataset to improve the quality of the $3$ D scene reconstruction.

Figure 2: Renders of \(3\) DGS models trained under different view selection metrics. Images are from the evaluation set and not seen during training. We find that our coverage metric is at least as good as random and superior to FisherRF. We highlight lower coverage regions in the random baseline in red . Additional visual comparisons and visualization of the view selection process can be found on our webpage and in Appendix D . — Figure 2: Renders of $3$ DGS models trained under different view selection metrics. Images are from the evaluation set and not seen during training. We find that our coverage metric is at least as good as random and superior to FisherRF. We highlight lower coverage regions in the random baseline in red . Additional visual comparisons and visualization of the view selection process can be found on our webpage and in Appendix D .

Figure 3: Image-based metrics (PSNR/SSIM/LPIPS) across several scenes for five view selection methods: Bayes’ Rays , FisherRF , Random , COVER , and an infeasible oracle (black) trained with all training images.

Dataset	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Tanks & Temples	Bayes’ Rays	15.03	0.43	0.62
	FisherRF	20.26	0.71	0.23
	Random	21.12	0.75	0.20
	CONVERGE (Ours)	21.52	0.76	0.19
Mip-NeRF	Bayes’ Rays	18.81	0.45	0.63
	FisherRF	24.92	0.78	0.19

Table 1: Image metrics of different view-selection methods at 30K steps across different datasets.

Setting	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
All	Splatfacto	24.83	0.79	0.16
Embodied	Random	22.48	0.71	0.23
	Fisher-RF	22.27	0.70	0.24
	CONVERGE (Ours)	23.21	0.73	0.20
Sparse	Random	22.81	0.72	0.21
	Fisher-RF	21.74	0.68	0.26

Table 2: Image metrics across different view-selection settings at 30K steps, averaged over all scenes. Splatfacto, with access to all views, serves as the infeasible upper-bound.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Scene	Method	PSNR $\uparrow$	$\Delta$ PSNR	SSIM $\uparrow$	$\Delta$ SSIM	LPIPS $\downarrow$	$\Delta$ LPIPS
Ignatius	Bayes’ Rays	16.71	-0.24	0.43	-0.12	0.56	0.24
	FisherRF	18.01	-0.56	0.59	-0.03	0.31	0.02
	Random	18.45	–	0.64	–	0.28	–
	CONVERGE (Ours)	19.40	0.45	0.67	0.01	0.27	-0.01
Train	Bayes’ Rays	12.70	-1.85	0.45	-0.07	0.73	0.28
	FisherRF	15.03	-0.77	0.57	-0.04	0.39	0.04

Appendix

Table 3: Image metrics across several view selection metrics for the Tanks and Temples dataset after

10

K gradient steps. Two scenes are visualized, while averages are computed across the entire dataset.

Scene	Method	PSNR $\uparrow$	$\Delta$ PSNR	SSIM $\uparrow$	$\Delta$ SSIM	LPIPS $\downarrow$	$\Delta$ LPIPS
Shiny	Bayes’ Rays	13.08	-1.43	0.31	-0.09	0.88	0.29
	FisherRF	13.69	-1.35	0.38	-0.05	0.58	0.07
	Random	15.47	–	0.46	–	0.47	–
	CONVERGE (Ours)	16.49	0.76	0.49	0.02	0.43	-0.03
Space	Bayes’ Rays	11.70	-1.74	0.20	-0.13	0.85	0.34
	FisherRF	13.94	-0.55	0.39	-0.03	0.46	0.03

Appendix

Table 4: Image metrics across several view selection metrics for the custom dataset after

10

K gradient steps. Two scenes are visualized, while averages are computed across the entire dataset.

Scene	Method	PSNR $\uparrow$	$\Delta$ PSNR	SSIM $\uparrow$	$\Delta$ SSIM	LPIPS $\downarrow$	$\Delta$ LPIPS
Counter	Bayes’ Rays	19.74	-2.51	0.61	-0.15	0.55	0.30
	FisherRF	22.22	-1.66	0.79	-0.04	0.25	0.04
	Random	24.06	–	0.83	–	0.22	–
	CONVERGE (Ours)	25.24	0.71	0.86	0.02	0.18	-0.02
Garden	Bayes’ Rays	17.62	-5.32	0.31	-0.31	0.77	0.51
	FisherRF	23.73	-0.95	0.72	-0.04	0.21	0.03

Appendix

Table 5: Image metrics across several view selection metrics for the MipNeRF360 dataset after

10

K gradient steps. Two scenes are visualized, while averages are computed across the entire dataset.

Scene	Method	PSNR $\uparrow$	$\Delta$ PSNR	SSIM $\uparrow$	$\Delta$ SSIM	LPIPS $\downarrow$	$\Delta$ LPIPS
Ignatius	Bayes’ Rays	17.10	-2.06	0.46	-0.20	0.45	0.25
	FisherRF	20.95	-0.28	0.71	-0.03	0.22	0.02
	Random	20.89	–	0.72	–	0.21	–
	CONVERGE (Ours)	21.53	0.68	0.73	0.02	0.20	-0.01
Train	Bayes’ Rays	13.04	-4.21	0.46	-0.20	0.70	0.41
	FisherRF	18.93	-1.34	0.72	-0.06	0.25	0.07

Appendix

Table 6: Image metrics across several view selection metrics for the Tanks and Temples dataset after

30

K gradient steps. Two scenes are visualized, while averages are computed across the entire dataset.

Scene	Method	PSNR $\uparrow$	$\Delta$ PSNR	SSIM $\uparrow$	$\Delta$ SSIM	LPIPS $\downarrow$	$\Delta$ LPIPS
Shiny	Bayes’ Rays	13.38	-3.08	0.30	-0.17	0.86	0.42
	FisherRF	16.97	-1.59	0.52	-0.06	0.41	0.09
	Random	18.27	–	0.56	–	0.33	–
	CONVERGE (Ours)	18.41	0.59	0.58	0.02	0.31	-0.03
Space	Bayes’ Rays	11.76	-3.85	0.19	-0.27	0.83	0.46
	FisherRF	18.16	-0.28	0.58	-0.02	0.29	0.03

Appendix

Table 7: Image metrics across several view selection metrics for the custom dataset after

30

K gradient steps. Two scenes are visualized, while averages are computed across the entire dataset.

Scene	Method	PSNR $\uparrow$	$\Delta$ PSNR	SSIM $\uparrow$	$\Delta$ SSIM	LPIPS $\downarrow$	$\Delta$ LPIPS
Counter	Bayes’ Rays	20.39	-4.90	0.62	-0.22	0.48	0.32
	FisherRF	26.35	-1.73	0.87	-0.04	0.17	0.04
	Random	27.51	–	0.89	–	0.15	–
	CONVERGE (Ours)	27.96	0.79	0.90	0.02	0.14	-0.02
Garden	Bayes’ Rays	19.28	-7.33	0.33	-0.44	0.75	0.59
	FisherRF	27.77	-1.01	0.85	-0.04	0.12	0.03

Appendix

Table 8: Image metrics across several view selection metrics for the MipNeRF360 dataset after

30

K gradient steps. Two scenes are visualized, while averages are computed across the entire dataset.

Setting	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Embodied + Sparse	Equation 17	21.95	0.695	0.251
	Gradient	21.77	0.701	0.275
	Coverage	22.39	0.707	0.244

Appendix

Table 9: Image metrics across different view-selection methods at 30K steps, averaged over all scenes.

Figure 4: Comparison of different view metrics in the space scene.

Figure 5: Comparison of different view metrics in the chair scene.

Explore similar work

A Bayesian Approach for Task-Specific Next-Best-View Selection with Uncertain Geometry
May 6, 2026Jingsen Zhu, Silvia Sellán, Alexander TereninNext-Best-View Planning 3D Reconstruction
Cornell University, USA · Columbia University, USA
GO-PRE: Goal-Oriented Next-Best-View Selection via Predictive Rendering Entropy for Active 3D Reconstruction
Jul 31, 2026Yan Song, Zhihao Li, Chenglong Li +3Next-Best-View Planning 3D Reconstruction
College of Intelligent Robotics and Advanced Manufacturing, Fudan University, Shanghai, China · School of Software, Shandong University, Jinan, China · School of Data Science and Engineering, East China Normal University, Shanghai, China
Matisse: Evidence-Space Reasoning for Active 3D Reconstruction
Sep 30, 2026Xihang Yu, Kaichen Zhou, Lorenzo Shaikewitz +4Next-Best-View Planning 3D Reconstruction
Massachusetts Institute of Technology · National University of Singapore

May 6, 2026J/K move · Enter open · S save

A Bayesian Approach for Task-Specific Next-Best-View Selection with Uncertain Geometry

Jingsen Zhu, Silvia Sellán, Alexander Terenin

Cornell University, USA · Columbia University, USA

Next-Best-View Planning 3D Reconstruction Bayesian Decision Theory Active View Selection Decision-Making under Uncertainty

Open paper AskPDF arXiv

We develop a framework for task-specific active next-best-view selection in 3D reconstruction from point clouds, by casting the problem in the language of Bayesian decision theory. Our framework works by (a) placing a prior distribution over the space of implicit surfaces, (b) using recently-developed stochastic surface reconstruction methods to calculate the resulting posterior distribution, then (c) using the posterior distribution to carefully reason about which view to scan next. This enables us to perform camera selection in a manner that is directly optimized for the intended use of the reconstructed data - meaning, we reduce uncertainty only in those regions that make a difference in the task at hand, as opposed to prior approaches that reduce it uniformly across space. We evaluate our method across three distinct downstream tasks: semantic classification, segmentation, and PDE-guided physics simulation. Experimental results demonstrate that our framework achieves superior task performance with fewer views compared to commonly used baselines and prior general uncertainty-reduction techniques.