Benchmarking Efficient & Effective Camera Pose Estimation Strategies for Novel View Synthesis

Jhacson Meza; Martin R. Oswald; Torsten Sattler

cs.CVMar 20, 2026

Benchmarking Efficient & Effective Camera Pose Estimation Strategies for Novel View Synthesis

Authors: Jhacson Meza, Martin R. Oswald, Torsten Sattler

Organizations: Faculty of Electrical Engineering, Czech Technical University in Prague · Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague · University of Amsterdam

Abstract

Novel view synthesis (NVS) approaches such as NeRFs or 3DGS can produce photo-realistic 3D scene representation from a set of images with known extrinsic and intrinsic parameters. The necessary camera poses and calibrations are typically obtained from the images via Structure-from-Motion (SfM). Classical SfM approaches rely on local feature matches between the images to estimate both the poses and a sparse 3D model of the scene, using bundle adjustment to refine initial pose, intrinsics, and geometry estimates. In order to increase run-time efficiency, recent SfM systems forgo optimization via bundle adjustment. Instead, they train feed-forward (transformer-based) neural networks to directly regress camera parameters and the 3D structure. While orders of magnitude more efficient, such recent works produce significantly less accurate estimates. To stimulate research on developing SfM approaches that are both efficient and effective, this paper develops a benchmark focused on SfM for novel view synthesis. Using existing datasets and two simple strategies for making the reconstruction process more efficient, we show that: (1) simply using fewer features already significantly accelerates classical SfM methods while maintaining high pose accuracy. (2) using feed-forward networks to obtain initial estimates and refining them using classical SfM techniques leads to the best efficiency-effectiveness trade-off. We will make our benchmark and code publicly available.

Figures & tables

Figure 1: Average absolute camera pose accuracy vs reconstruction time across the 25 scenes from the ETH3D MVS dataset using LightGlue and MNN matchers and COLMAP and GLOMAP mappers. Runtimes are average across all the scenes. For the same matcher, both time and accuracy scales are fixed to allow a direct comparison between the mappers.

Figure 2: PSNR vs. reconstruction runtime results using the LightGlue matcher for Mip-NeRF 360, Tanks and Temples, and Zip-NeRF datasets. Missing views for the average PSNR computation for a scene are penalized with a value of 0. Scenes that failed to reconstruct or render are also penalize with 0 for the per scene aggregation.

Figure 3: PSNR using NeRF vs. reconstruction runtime results with LightGLue matcher for Mip-NeRF 360, Tanks and Temples, and Zip-NeRF datasets.

	<1° and <5cm	<5° and <10cm	Time (s)
Fast3R	1.22%	14.92%	1.08
VGGT	10.80%	44.32%	2.12
$\pi^{3}$	22.38%	62.36%	2.71
Depth Anything 3	47.22%	78.40%	2.45
COLMAP-LG-SP (256)	48.33%	49.44%	65.48
COLMAP-LG-SP (512)	56.68%	62.14%	107.60

Table 1: Average absolute camera pose accuracy across all the scenes from ETH3D MVS dataset for the feed-forward SfM methods and traditional SfM pipelines. We highlight the best , second , and third values.

	Mip-NeRF 360			Tanks and Temples			Zip-NeRF
	Mip-Splat.	Nerfacto	Time (min)	Mip-Splat.	Nerfacto	Time (min)	Mip-Splat.	Nerfacto	Time (min)
Fast3R	15.06	15.12	0.26	12.61	13.38	0.33	OOM	OOM	-
VGGT	2.22	1.86	-	1.78	1.96	-	OOM	OOM	-
$\pi^{3}$	16.03	14.30	0.69	11.42	9.71	0.95	OOM	OOM	-
Depth Anything 3	17.50	15.55	0.34	11.93	12.28	0.30	OOM	OOM	-
COLMAP-LG-SP (256)	24.07	22.15	12.43	20.37	18.86	14.49	20.66	20.01	50.11

Table 2: Average PSNR results on multiple datasets for feed-forward SfM methods and different configuration for traditional SfM.

	Standard COLMAP reconstruction										Feed-forward recons
	512		1024		2048		4096		8192		DA3		Refined DA3
Scene	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time
bicycle	24.02	1.13	24.73	1.39	25.37	2.31	25.59	3.95	25.62	9.33	18.54	0.42	25.84	2.83
bonsai	31.16	1.80	31.59	3.32	24.96	11.51	32.42	12.24	32.40	24.91	21.10	0.69	-	14.45
counter	28.70	1.16	28.94	1.55	29.09	3.77	28.89	7.32	29.16	15.21	22.89	0.48	29.21	5.86
flowers	1.12	0.61	10.30	1.03	21.16	1.84	21.27	2.69	21.68	7.85	16.86	0.28	22.25	1.63

Table 3: Rendering and reconstruction times for the refinement of Depth-Anything-3 camera poses using 8192 features for the Mip-NeRF 360 dataset. We also show PSNR and times for standard COLMAP reconstruction using different number of features. Times are in minutes.

	Standard COLMAP reconstruction										Feed-forward recons
	512		1024		2048		4096		8192		$\pi^{3}$		Refined $\pi^{3}$
Scene	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time
auditorium	23.03	2.67	23.28	2.18	23.32	4.64	23.27	8.98	23.31	6.06	16.55	1.21	22.65	5.54
ballroom	21.88	2.39	19.69	5.56	19.78	9.46	23.11	18.60	17.13	18.28	12.43	1.40	16.15	19.36
barn	27.44	2.58	27.66	4.44	27.86	8.91	27.96	13.51	27.95	13.56	OOM	OOM	OOM	OOM
caterpillar	23.11	2.47	23.04	3.78	23.31	7.60	23.38	12.76	23.39	16.03	15.04	1.91	23.54	13.20

Table 4: Rendering and time results on FFN refinements using 8192 feats and Tanks & Temples dataset. Times are in minutes.

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Average runtimes for each stage of the Structure-from-Motion pipeline for the Mip-NeRF 360 dataset.

Figure 5: Average runtimes for each stage of the Structure-from-Motion pipeline for the Tanks and Temples dataset.

Figure 6: Average runtimes for each stage of the Structure-from-Motion pipeline for the Zip-NeRF dataset.

Figure 7: PSNR vs. reconstruction runtime results using the MNN matcher for Mip-NeRF 360, Tanks and Temples, and Zip-NeRF datasets.

Figure 8: SSIM and runtime results for the Mip-NeRF 360 dataset.

Figure 9: LPIPS and runtime results for the Mip-NeRF 360 dataset.

Figure 10: SSIM and runtime results for the Tanks and Temples dataset.

Figure 11: LPIPS and runtime results for the Tanks and Temples dataset.

Figure 12: SSIM and runtime results for the Zip-NeRF dataset.

Figure 13: LPIPS and runtime results for the Zip-NeRF dataset.

Figure 14: Renderings for different number of keypoints using COLMAP’s feature extractor and matcher.

Features	Detector	bicycle	bonsai	counter	flowers	garden	kitchen	room	stump	treehill
128	ALIKED	—	30.61	—	—	26.41	21.39	30.53	—	2.74
	DISK	0.68	13.57	27.13	—	26.01	22.41	13.40	—	—
	DoG-HardNet	7.03	22.71	9.86	—	24.58	17.36	13.90	4.56	—
	SIFT	—	22.82	9.49	—	24.93	17.12	11.98	4.61	—
	SuperPoint	—	30.60	—	—	25.94	28.78	2.08	—	—
	XFeat	—	—	—	—	—	—	—	—	—

Appendix

Table 5: Per-scene PSNR results on Mip-NeRF 360 dataset using COLMAP mapper and LG matcher. Cells with dashes represent cases where the sparse reconstruction process failed. We highlight the best , second , and third values for each scene.

Features	Detector	bicycle	bonsai	counter	flowers	garden	kitchen	room	stump	treehill
128	ALIKED	1.51	30.61	25.85	—	26.14	25.15	14.46	—	4.92
	DISK	4.65	—	25.77	2.03	25.44	15.21	12.75	1.16	1.33
	DoG-HardNet	6.34	22.76	7.31	—	22.00	13.11	25.19	6.18	0.79
	SIFT	6.33	17.28	6.59	—	24.54	10.42	17.70	6.18	0.81
	SuperPoint	4.81	26.09	13.32	0.98	24.36	15.37	29.92	1.05	2.51
	XFeat	—	8.43	1.32	—	—	19.74	18.38	—	—

Appendix

Table 6: Per-scene PSNR results on Mip-NeRF 360 dataset using GLOMAP mapper and LG matcher

Features	Detector	auditorium	ballroom	barn	caterpillar	church	courthouse	courtroom	family	francis	horse	ignatius	lighthouse	m60	meetingroom	museum	palace	panther	playground	temple	train	truck
128	ALIKED	22.17	13.26	19.26	22.69	20.36	20.61	14.61	16.68	13.83	20.32	—	19.34	26.82	—	11.95	10.54	27.69	13.41	12.11	14.76	9.39
	DISK	—	7.18	12.83	13.95	0.29	18.93	7.70	1.39	17.58	0.73	4.95	14.15	14.35	10.02	9.78	10.30	20.22	4.92	7.95	12.72	3.03
	DoG-HardNet	—	0.50	2.03	15.21	14.48	7.39	8.92	—	16.35	—	8.39	5.37	0.91	6.53	4.21	2.91	0.65	—	0.47	0.75	—
	SIFT	—	0.51	1.80	14.82	14.04	4.51	8.97	0.62	16.36	—	1.62	—	0.28	—	3.15	3.44	0.25	2.07	—	0.75	—
	SuperPoint	20.03	—	25.48	—	10.76	20.17	—	9.18	14.29	—	—	11.69	—	13.02	3.35	6.35	—	—	0.97	0.72	—
	XFeat	20.40	—	24.09	—	14.60	15.56	—	—	13.43	—	—	11.00	—	—	—	—	—	—	3.80	—	—

Appendix

Table 7: Per-scene PSNR results on Tanks and Temples dataset using COLMAP mapper and LG matcher

Features	Detector	auditorium	ballroom	barn	caterpillar	church	courthouse	courtroom	family	francis	horse	ignatius	lighthouse	m60	meetingroom	museum	palace	panther	playground	temple	train	truck
128	ALIKED	—	20.45	25.90	12.38	—	14.52	15.57	—	14.53	21.85	13.53	17.86	26.83	11.90	13.06	10.69	27.65	—	8.84	—	—
	DISK	—	11.40	12.37	13.77	—	13.82	14.77	14.15	13.13	17.48	15.22	12.77	18.59	—	12.25	—	23.41	—	9.35	14.29	17.50
	DoG-HardNet	7.47	—	3.69	—	—	6.67	—	4.77	24.27	6.48	—	6.64	6.67	7.38	3.30	2.73	17.79	6.38	0.80	6.31	2.81
	SIFT	5.68	8.69	5.52	9.71	10.50	11.52	10.88	2.92	16.69	3.62	11.48	—	4.67	5.09	3.49	3.01	6.00	—	0.98	8.12	4.37
	SuperPoint	16.45	3.87	15.30	20.76	8.16	19.76	4.47	18.27	—	—	12.58	13.61	8.00	15.59	4.46	4.64	24.14	11.14	3.83	13.99	—
	XFeat	10.30	10.21	13.69	6.59	—	—	3.55	6.11	12.34	8.95	—	11.96	1.87	8.00	3.81	6.48	7.34	1.36	5.36	3.86	—

Appendix

Table 8: Per-scene PSNR results on Tanks and Temples dataset using GLOMAP mapper and LG matcher

Table 23

	Standard COLMAP reconstruction										Feed-forward recons
	512		1024		2048		4096		8192		DA3		Refined DA3		$\pi^{3}$		Refined $\pi^{3}$		Fast3R		Refined Fast3R		VGGT		Refined VGGT
Scene	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time
bicycle	24.02	1.13	24.73	1.39	25.37	2.31	25.59	3.95	25.62	9.33	18.54	0.42	25.84	2.83	14.96	0.60	-	5.71	14.85	0.28	14.11	5.86	OOM	OOM	OOM	OOM
bonsai	31.16	1.80	31.59	3.32	24.96	11.51	32.42	12.24	32.40	24.91	21.10	0.69	-	14.45	18.66	1.11	32.76	12.75	15.54	0.42	20.68	22.23	OOM	OOM	OOM	OOM
counter	28.70	1.16	28.94	1.55	29.09	3.77	28.89	7.32	29.16	15.21	22.89	0.48	29.21	5.86	16.98	0.77	29.20	8.41	16.08	0.29	20.82	10.94	OOM	OOM	OOM	OOM
flowers	1.12	0.61	10.30	1.03	21.16	1.84	21.27	2.69	21.68	7.85	16.86	0.28	22.25	1.63	11.51	0.43	18.28	3.57	10.09	0.16	11.93	1.31	OOM	OOM	OOM	OOM

Appendix

Table 11: Rendering and time results of the feed-forward refinements using 8192 feats on the MipNeRF 360 dataset. Times are in minutes.

	Standard COLMAP reconstruction										Feed-forward recons
	512		1024		2048		4096		8192		DA3		Refined DA3		$\pi^{3}$		Refined $\pi^{3}$		Fast3R		Refined Fast3R		VGGT		Refined VGGT
Scene	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time
bicycle	24.02	1.13	24.73	1.39	25.37	2.31	25.59	3.95	25.62	9.33	18.54	0.42	25.62	2.57	14.96	0.60	18.38	3.54	14.85	0.28	13.00	3.47	OOM	OOM	OOM	OOM
bonsai	31.16	1.80	31.59	3.32	24.96	11.51	32.42	12.24	32.40	24.91	21.10	0.69	-	8.41	18.66	1.11	32.73	8.36	15.54	0.42	16.73	15.87	OOM	OOM	OOM	OOM
counter	28.70	1.16	28.94	1.55	29.09	3.77	28.89	7.32	29.16	15.21	22.89	0.48	29.20	3.82	16.98	0.77	29.20	4.68	16.08	0.29	-	0.29	OOM	OOM	OOM	OOM
flowers	1.12	0.61	10.30	1.03	21.16	1.84	21.27	2.69	21.68	7.85	16.86	0.28	22.07	1.40	11.51	0.43	16.93	2.26	10.09	0.16	-	0.16	OOM	OOM	OOM	OOM

Appendix

Table 12: Rendering and time results of the feed-forward refinements using 4096 feats on the MipNeRF 360 dataset. Times are in minutes.

	Standard COLMAP reconstruction										Feed-forward recons
	512		1024		2048		4096		8192		DA3		Refined DA3		$\pi^{3}$		Refined $\pi^{3}$		Fast3R		Refined Fast3R		VGGT		Refined VGGT
Scene	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time
bicycle	24.02	1.13	24.73	1.39	25.37	2.31	25.59	3.95	25.62	9.33	18.54	0.42	25.24	1.80	14.96	0.60	24.82	2.63	14.85	0.28	14.87	2.52	OOM	OOM	OOM	OOM
bonsai	31.16	1.80	31.59	3.32	24.96	11.51	32.42	12.24	32.40	24.91	21.10	0.69	-	5.87	18.66	1.11	32.15	5.29	15.54	0.42	18.20	6.85	OOM	OOM	OOM	OOM
counter	28.70	1.16	28.94	1.55	29.09	3.77	28.89	7.32	29.16	15.21	22.89	0.48	29.03	2.36	16.98	0.77	29.04	3.83	16.08	0.29	-	0.29	OOM	OOM	OOM	OOM
flowers	1.12	0.61	10.30	1.03	21.16	1.84	21.27	2.69	21.68	7.85	16.86	0.28	21.71	1.04	11.51	0.43	17.45	1.63	10.09	0.16	-	0.16	OOM	OOM	OOM	OOM

Appendix

Table 13: Rendering and time results of the feed-forward refinements using 2048 feats on the MipNeRF 360 dataset. Times are in minutes.

	Standard COLMAP reconstruction										Feed-forward recons
	512		1024		2048		4096		8192		DA3		Refined DA3		$\pi^{3}$		Refined $\pi^{3}$		Fast3R		Refined Fast3R		VGGT		Refined VGGT
Scene	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time
auditorium	23.03	2.67	23.28	2.18	23.32	4.64	23.27	8.98	23.31	6.06	21.47	0.52	-	7.14	16.55	1.21	22.65	5.54	17.90	0.30	16.61	6.53	OOM	OOM	OOM	OOM
ballroom	21.88	2.39	19.69	5.56	19.78	9.46	23.11	18.60	17.13	18.28	13.08	0.60	15.67	17.26	12.43	1.40	16.15	19.36	11.97	0.34	13.17	17.85	OOM	OOM	OOM	OOM
barn	27.44	2.58	27.66	4.44	27.86	8.91	27.96	13.51	27.95	13.56	OOM	OOM	OOM	OOM	OOM	OOM	OOM	OOM	10.28	0.53	-	0.53	OOM	OOM	OOM	OOM
caterpillar	23.11	2.47	23.04	3.78	23.31	7.60	23.38	12.76	23.39	16.03	OOM	OOM	OOM	OOM	15.04	1.91	23.54	13.20	12.29	0.47	13.04	18.20	OOM	OOM	OOM	OOM

Appendix

Table 14: Rendering and time results of the feed-forward refinements using 8192 feats on the Tanks & Temples. Times are in minutes.

	Standard COLMAP reconstruction										Feed-forward recons
	512		1024		2048		4096		8192		DA3		Refined DA3		$\pi^{3}$		Refined $\pi^{3}$		Fast3R		Refined Fast3R		VGGT		Refined VGGT
Scene	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time
auditorium	23.03	2.67	23.28	2.18	23.32	4.64	23.27	8.98	23.31	6.06	21.47	0.52	-	6.84	16.55	1.21	22.16	6.25	17.90	0.30	14.52	6.60	OOM	OOM	OOM	OOM
ballroom	21.88	2.39	19.69	5.56	19.78	9.46	23.11	18.60	17.13	18.28	13.08	0.60	15.60	15.12	12.43	1.40	16.14	15.48	11.97	0.34	13.39	15.19	OOM	OOM	OOM	OOM
barn	27.44	2.58	27.66	4.44	27.86	8.91	27.96	13.51	27.95	13.56	OOM	OOM	OOM	OOM	OOM	OOM	OOM	OOM	10.28	0.53	-	0.53	OOM	OOM	OOM	OOM
caterpillar	23.11	2.47	23.04	3.78	23.31	7.60	23.38	12.76	23.39	16.03	OOM	OOM	OOM	OOM	15.04	1.91	23.37	18.20	12.29	0.47	13.15	13.35	OOM	OOM	OOM	OOM

Appendix

Table 15: Rendering and time results of the feed-forward refinements using 4096 feats on the Tanks & Temples. Times are in minutes.

	Standard COLMAP reconstruction										Feed-forward recons
	512		1024		2048		4096		8192		DA3		Refined DA3		$\pi^{3}$		Refined $\pi^{3}$		Fast3R		Refined Fast3R		VGGT		Refined VGGT
Scene	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time	PSNR	Time
auditorium	23.03	2.67	23.28	2.18	23.32	4.64	23.27	8.98	23.31	6.06	21.47	0.52	22.64	6.14	16.55	1.21	22.36	6.73	17.90	0.30	18.72	5.17	OOM	OOM	OOM	OOM
ballroom	21.88	2.39	19.69	5.56	19.78	9.46	23.11	18.60	17.13	18.28	13.08	0.60	15.33	8.73	12.43	1.40	15.68	9.66	11.97	0.34	12.89	10.46	OOM	OOM	OOM	OOM
barn	27.44	2.58	27.66	4.44	27.86	8.91	27.96	13.51	27.95	13.56	OOM	OOM	OOM	OOM	OOM	OOM	OOM	OOM	10.28	0.53	-	0.53	OOM	OOM	OOM	OOM
caterpillar	23.11	2.47	23.04	3.78	23.31	7.60	23.38	12.76	23.39	16.03	OOM	OOM	OOM	OOM	15.04	1.91	23.34	10.80	12.29	0.47	12.89	13.77	OOM	OOM	OOM	OOM

Appendix

Table 16: Rendering and time results of the feed-forward refinements using 2048 feats on the Tanks & Temples. Times are in minutes.

Detector	Scene	128				256				512				1024				2048				4096				8192
Detector	Scene	2D points	Matches	3D points	Reg. views	2D points	Matches	3D points	Reg. views	2D points	Matches	3D points	Reg. views	2D points	Matches	3D points	Reg. views	2D points	Matches	3D points	Reg. views	2D points	Matches	3D points	Reg. views	2D points	Matches	3D points	Reg. views
ALIKED	bicycle	24.8K	0.2M	38	2	49.7K	0.4M	6.4K	193	99.3K	0.7M	1.5K	194	0.2M	1.1M	3.0K	194	0.4M	2.1M	5.8K	194	0.8M	4.0M	1.1K	194	1.0M	4.7M	1.4K	194
	bonsai	37.4K	0.5M	4.5K	282	74.8K	1.1M	8.7K	292	0.1M	1.9M	1.5K	292	0.3M	3.8M	2.3K	292	0.6M	7.8M	3.7K	292	0.8M	10.8M	4.8K	292	0.8M	10.8M	4.8K	292
	counter	30.7K	0.3M	35	2	61.4K	0.7M	8.6K	240	0.1M	1.3M	1.5K	240	0.2M	2.5M	2.7K	240	0.5M	4.9M	4.8K	240	0.6M	6.2M	5.9K	240	0.6M	6.2M	5.9K	240
	flowers	22.1K	0.1M	28	2	44.3K	0.3M	37	2	88.6K	0.6M	1.3K	173	0.2M	1.0M	2.7K	173	0.4M	1.7M	5.2K	173	0.7M	3.0M	1.0K	173	0.8M	3.4M	1.2K	173
	garden	23.7K	0.2M	3.1K	185	47.4K	0.5M	6.4K	185	94.7K	0.9M	1.2K	185	0.2M	1.9M	2.1K	185	0.4M	3.9M	3.8K	185	0.8M	8.0M	7.1K	185	0.8M	8.7M	7.7K	185

Appendix

Table 17: 3D reconstruction stats on MipNeRF 360 dataset using COLMAP mapper and LightGlue matcher

Detector	Scene	128				256				512				1024				2048				4096				8192
Detector	Scene	2D points	Matches	3D points	Reg. views	2D points	Matches	3D points	Reg. views	2D points	Matches	3D points	Reg. views	2D points	Matches	3D points	Reg. views	2D points	Matches	3D points	Reg. views	2D points	Matches	3D points	Reg. views	2D points	Matches	3D points	Reg. views
ALIKED	bicycle	24.8K	0.6M	35	2	49.7K	1.0M	6.1K	188	99.3K	2.0M	1.5K	193	0.2M	3.7M	2.9K	193	0.4M	6.8M	5.7K	194	0.8M	11.3M	1.1K	194	1.0M	12.4M	1.3K	194
	bonsai	37.4K	1.3M	5.0K	292	74.8K	2.4M	9.2K	292	0.1M	4.3M	1.8K	292	0.3M	7.7M	2.7K	292	0.6M	13.1M	4.5K	292	0.8M	16.1M	6.0K	292	0.8M	16.1M	6.0K	292
	counter	30.7K	0.9M	276	13	61.4K	1.6M	9.4K	240	0.1M	3.2M	1.7K	236	0.2M	5.9M	3.2K	240	0.5M	9.5M	5.7K	240	0.6M	10.7M	7.2K	240	0.6M	10.7M	7.3K	240
	flowers	22.1K	0.4M	41	2	44.3K	0.8M	25	2	88.6K	1.6M	1.0K	139	0.2M	3.0M	2.6K	173	0.4M	5.6M	5.1K	173	0.7M	9.7M	9.8K	173	0.8M	10.9M	1.1K	173
	garden	23.7K	0.6M	3.3K	185	47.4K	1.1M	6.7K	185	94.7K	2.1M	1.3K	185	0.2M	4.1M	2.3K	185	0.4M	7.6M	4.1K	185	0.8M	13.3M	7.6K	185	0.8M	14.1M	8.2K	185

Appendix

Table 18: 3D reconstruction stats on MipNeRF 360 dataset using COLMAP mapper and Mutual Nearest Neighbors matcher

Explore similar work

Oct 7, 2026

LVSPM: Long Sequence View Synthesis and Pose Estimation Model

We present LVSPM, a generalizable model that jointly estimates camera poses and synthesizes novel views from uncalibrated image collections. Trained with only RGB images and pose supervision, LVSPM avoids dense 3D ground truth and employs test-time training (TTT) layers to scale seamlessly to hundreds of input views. On RealEstate10k, Co3Dv2, and DL3DV, LVSPM surpasses VGGT in pose estimation across 16-256 views, with especially large margins at strict thresholds. For novel view synthesis under a practical protocol where more views cover larger scenes, LVSPM achieves state-of-the-art pose-free quality---surpassing even pose-dependent models in PSNR---and still maintains high quality as scene scale grows, while baselines collapse. The code is available at https://burningdust21.github.io/Projects/LVSPM .

Xi Chen, Yachi Zhang, Linghao Chen +4

UC San Diego · Sudo AI GmbH

Novel View Synthesis Camera Pose Estimation

Ask PDF

Sep 8, 2026

Learning Global Camera Poses from Noisy View-Graphs for Structure from Motion

Camera pose estimation is a key step in 3D reconstruction and view-synthesis pipelines. We present a deep, global Structure-from-Motion framework based on learned view-graph aggregation. Our method employs a permutation-equivariant, edge-conditioned graph neural network that takes noisy pairwise relative poses as input and outputs globally consistent camera extrinsics. The network is trained without ground-truth supervision, relying solely on a relative-pose consistency objective. This is followed by 3D point triangulation and robust bundle adjustment. Our approach is efficient, scalable to more than a thousand images, and robust to graph density. We evaluate our method on MegaDepth, 1DSfM, Strecha, and BlendedMVS. These experiments demonstrate that our method achieves superior rotation and translation accuracy compared to deep track-centric methods while registering more images across many scenes, and competitive results compared to state-of-the-art classical pipelines, while being much faster.

Fadi Khatib, Meirav Galun, Ronen Basri

Weizmann Institute of Science

Relative Pose Estimation Camera Pose Estimation

Ask PDF

Aug 22, 2026

ORBIT++: Benchmarking SfM in the Wild with 360° Video

Structure-from-Motion (SfM) is a cornerstone of 3D perception, yet current methods often fail when applied to complex videos involving challenging camera motions or dynamic scenes. Compounding the problem, the field lacks reliable ground-truth benchmarks for such difficult scenarios, making it hard to gauge real-world progress or to pinpoint where improvements are most needed. To address this gap, we introduce a new benchmark for evaluating camera pose estimation. Our key insight is to leverage online panoramic 360° video as a source of data from which to construct challenging clips, while still enabling robust ground-truth trajectory recovery. The panoramic nature of these videos provides richer visual context for tracking camera motion, even when parts of the view are affected by blur, motion, or dynamic objects. After tracking camera motion across full 360° videos, we crop and reproject selected portions to generate perspective-view clips that serve as our benchmark, called ORBIT. Experiments show that COLMAP, as well as recent optimization-based and feed-forward SfM methods struggle to accurately estimate camera poses on our benchmark. Hence, ORBIT provides a valuable testbed where researchers can meaningfully measure progress on truly challenging, real-world SfM problems.

Sara Sabour, Linyi Jin, Richard Tucker +9

Google DeepMind · Simon Fraser University · University of Toronto

Structure from Motion Camera Pose Estimation

Ask PDF