Privacy-Preserving Structureless Visual Localization via Image Obfuscation

Vojtech Panek; Patrik Beliansky; Zuzana Kukelova; Torsten Sattler

cs.CVApr 13, 2026

Privacy-Preserving Structureless Visual Localization via Image Obfuscation

Authors: Vojtech Panek, Patrik Beliansky, Zuzana Kukelova, Torsten Sattler

Organizations: Faculty of Electrical Engineering, Czech Technical University (CTU) in Prague · Czech Institute of Informatics, Robotics and Cybernetics, CTU in Prague · Visual Recognition Group, Faculty of Electrical Engineering, CTU in Prague

Abstract

Visual localization is the task of estimating the camera pose of an image relative to a scene representation. In practice, visual localization systems are often cloud-based. Naturally, this raises privacy concerns in terms of revealing private details through the images sent to the server or through the representations stored on the server. Privacy-preserving localization aims to avoid such leakage of private details. However, the resulting localization approaches are significantly more complex, slower, and less accurate than their non-privacy-preserving counterparts. In this paper, we consider structureless localization methods in the context of privacy preservation. Structureless methods represent the scene through a set of reference images with known camera poses and intrinsics. In contrast to existing methods proposing representations that are as privacy-preserving as possible, we study a simple image obfuscation approach based on common image operations, e.g., replacing RGB images with (semantic) segmentations. We show that existing structureless pipelines do not need any special adjustments, as modern feature matchers can match obfuscated images out of the box. The results are easy-to-implement pipelines that can ensure both the privacy of the query images and the scene representations. Detailed experiments on multiple datasets show that the resulting methods achieve state-of-the-art pose accuracy for privacy-preserving approaches.

Figures & tables

Figure 1: In this paper, we investigate privacy-preservation through image obfuscation in the context of visual localization. Rather than using the original images, we obfuscate the query and database photos. We show that this approach to privacy-preservation requires no adaption to existing pipelines and we investigate the impact on localization accuracy for various obfuscation schemes.

Panel of Figure 2 — Figure 2: Example segmentation masks: The top row shows the original image , SAM1 - fine masks , and SAM1 - fine borders . The bottom row shows Mask2Former segmentation masks: colored by semantic labels ( semantic ), colored randomly ( random ), segment borders ( borders ).

Panel of Figure 3 — Figure 3: Segmentation masks for (from top left): SAM2 coarse , SAM2 fine , SAM1 coarse , SAM1 fine .

Panel of Figure 4 — Figure 4: Example of text masking for SAM sampling. In the top row: original image and SAM1 fine mask . The bottom row: detected text regions and SAM1 fine mask with the point sampling for SAM inference limited to areas without text.

	day	night
original images	84.7 / 93.3 / 98.2	67.5 / 88.5 / 98.4
easy-anon - single	85.2 / 93.1 / 98.1	67.0 / 86.9 / 98.4
SAM1 - fine borders	74.0 / 84.5 / 93.9	60.2 / 81.7 / 95.8
LDP-FEAT [ 80 ]	76.1 / 85.0 / 90.4	24.6 / 29.8 / 34.0
Hloc [ 84 ] (SP+LG)	88.1 / 95.4 / 99.0	71.7 / 89.0 / 97.9

Table 1: Brief comparison on Aachen Day-Night v1.1 [ 110 , 87 , 89 ] using the top-20 reference images retrieved with EigenPlaces [ 8 ] . The first three rows use RoMa [ 33 ] matching and pose estimation with local triangulation. Hloc [ 84 ] (non-privacy-preserving, structure-based) uses its standard configuration with SuperPoint [ 23 ] features and LightGlue [ 59 ] matcher. We report the percentage of queries localized within error thresholds of (0.25 m, 2°) / (0.5 m, 5°) / (5 m, 10°).

	Aachen v1.1		Aachen v1.1		Bedroom		Liv. Room		Liv. Room		Office
	day		night		Table		Pillows		Sofa		Chairs
method	E5+1	LT	E5+1	LT	E5+1	LT	E5+1	LT	E5+1	LT	E5+1	LT
original images	82.4 / 93.9 / 99.2	84.7 / 93.3 / 98.2	70.7 / 89.0 / 98.4	67.5 / 88.5 / 98.4	93.5	96.5	94.6	95.9	98.5	97.7	89.4	90.3
easy-anon - single	82.8 / 93.8 / 99.2	85.2 / 93.1 / 98.1	72.8 / 87.4 / 98.4	67.0 / 86.9 / 98.4	94.2	95.8	94.6	96.4	98.5	9 8.5	88.5	88.5
easy-anon - inpaint	82.4 / 93.8 / 99.0	85.1 / 93.7 / 98.1	73.3 / 88.0 / 98.4	68.1 / 88.0 / 97.9	94.2	96.2	94.6	95.9	98.5	98.5	89.4	89.7
SAM2 - coarse masks	28.6 / 46.8 / 77.7	40.7 / 58.6 / 80.0	17.8 / 34.0 / 70.2	28.3 / 44.0 / 74.3	46.2	51.5	66.1	75.6	56.8	50.8	47.6	60.3

Table 2: Ablations on Aachen Day-Night v1.1 [ 110 , 87 , 89 ] and Remove360 [ 48 , 49 ] using the top-20 reference images retrieved with EigenPlaces [ 8 ] , RoMa [ 33 ] feature matching, and pose estimation with the E5+1 solver or via local triangulation (LT). We report the percentage of queries localized within error thresholds of (0.25 m, 2°) / (0.5 m, 5°) / (5 m, 10°) for Aachen and (5 cm, 5°) for Remove360.

		Great Court		King’s College		Old Hospital		Shop Façade		St Mary’s Church		avg. (4 scenes)		avg. (5 scenes)
		MPE	MOE	MPE	MOE	MPE	MOE	MPE	MOE	MPE	MOE	MPE	MOE	MPE	MOE
method		[m]	[°]	[m]	[°]	[m]	[°]	[m]	[°]	[m]	[°]	[m]	[°]	[m]	[°]
E5+1	original images	0.32	0.15	0.22	0.31	0.27	0.53	0.07	0.27	0.11	0.35	0.17	0.37	0.20	0.32
E5+1	SAM1 - fine borders	0.40	0.18	0.22	0.32	0.30	0.54	0.08	0.35	0.13	0.40	0.18	0.40	0.23	0.36
LT	original images	0.33	0.14	0.21	0.32	0.26	0.51	0.07	0.27	0.11	0.35	0.22	0.31	0.20	0.32
LT	SAM1 - fine borders	0.39	0.18	0.21	0.30	0.29	0.53	0.08	0.37	0.13	0.39	0.31	0.35	0.22	0.35

Table 3: Comparison to other privacy-preserving pipelines on Cambridge Landmarks [ 44 ] using top-10 retrieval with NetVLAD [ 4 ] (for a fair comparison with the baselines) and RoMa [ 33 ] matching. We are reporting median position (MPE) and orientation (MOE) errors (smaller is better). The average over 4 scenes excludes the Great Court scene.

		scene1			scene2a			scene3			scene4a			scene5			scene6			average
		MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.
method		[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]
E5+1	easy-anon - single	0.01	0.18	92.4	0.01	0.12	93.0	0.01	0.13	96.2	0.01	0.21	93.0	0.01	0.20	87.5	0.01	0.13	94.1	0.01	0.16	92.7
	easy-anon - inpaint	0.01	0.17	92.0	0.01	0.12	91.4	0.01	0.13	96.5	0.01	0.20	96.8	0.01	0.20	89.9	0.01	0.12	95.0	0.01	0.16	93.6
	SAM1 - fine masks	0.02	0.42	75.8	0.02	0.24	79.0	0.02	0.32	77.8	0.02	0.51	72.2	0.03	0.41	68.4	0.01	0.28	80.8	0.02	0.36	75.7
	SAM1 - fine borders	0.02	0.34	78.0	0.02	0.21	79.8	0.01	0.27	85.7	0.02	0.46	74.1	0.02	0.41	71.9	0.01	0.21	88.2	0.02	0.32	79.6

Table 4: Localization results on Indoor-6 [ 25 ] using the top-20 reference images retrieved with EigenPlaces [ 8 ] , RoMa [ 33 ] matching, and pose estimation with E5+1 and local triangulation (LT). The bottom part of the table contains results of other privacy-preserving methods. Reporting median position (MPE) and orientation (MOE) errors (smaller is better) and recall (rec.) at 5 cm, 5° pose error (higher is better).

obfuscation		filtered	day	night
original images		✗	77.9 / 90.5 / 98.2	64.9 / 88.5 / 98.4
SAM1	fine masks	✗	63.6 / 80.5 / 94.9	42.4 / 73.8 / 96.9
	fine masks	✓	63.1 / 79.0 / 94.2	47.1 / 76.4 / 96.3
	fine borders	✗	71.7 / 83.1 / 96.1	53.9 / 80.1 / 97.4
	fine borders	✓	70.1 / 84.2 / 95.5	53.4 / 78.5 / 96.9

Table 5: Ablation on SAM sampling with text masks. Aachen Day-Night v1.1 [ 110 , 87 , 89 ] using top-20 retrieval with EigenPlaces [ 8 ] , RoMa [ 33 ] matching and E5+1. Reporting recalls (higher is better) at the error thresholds of (0.25 m, 2°) / (0.5 m, 5°) / (5 m, 10°).

				Great		King’s		Old		Shop		St Mary’s		avg.		avg.
				Court		College		Hospital		Façade		Church		(4 scenes)		(5 scenes)
			ref. poses	MPE	MOE	MPE	MOE	MPE	MOE	MPE	MOE	MPE	MOE	MPE	MOE	MPE	MOE
method			ref. poses	[m]	[°]	[m]	[°]	[m]	[°]	[m]	[°]	[m]	[°]	[m]	[°]	[m]	[°]
SP+LG	E5+1	easy-anon - single	obf.	2.30	1.35	0.37	0.52	0.35	0.63	0.11	0.47	0.13	0.38	0.24	0.50	0.65	0.67
		easy-anon - single	g.t.	0.38	0.17	0.17	0.27	0.23	0.45	0.06	0.23	0.10	0.31	0.14	0.32	0.19	0.29
		SAM1 - fine borders	obf.	2.44	1.73	0.63	0.94	0.42	0.79	0.15	0.51	0.30	1.04	0.37	0.82	0.79	1.00

Table 6: "End-to-end" pipeline using reference poses from SfM on obfuscated images (obf.) or ground-truth reference poses (g.t.). Evaluated on Cambridge Landmarks [ 44 ] using top-10 NetVLAD [ 4 ] retrieval. We are reporting median position (MPE) and orientation (MOE) errors (smaller is better). The average over 4 scenes excludes the Great Court scene.

query		ref.		matching	day	night
images	orig.	images	orig.	RoMa	82.4 / 93.9 / 99.2	70.7 / 89.0 / 98.4
				SP+LG	78.8 / 92.2 / 98.7	65.4 / 83.2 / 97.4
				ALIKED+LG	79.4 / 92.8 / 98.9	67.0 / 81.2 / 97.9
				MASt3R	82.8 / 95.3 / 99.6	64.4 / 84.3 / 97.9
masks	SAM1 f.	masks	SAM1 f.	RoMa	63.6 / 80.5 / 94.9	42.4 / 73.8 / 96.9
				SP+LG	51.9 / 69.1 / 87.7	30.4 / 51.8 / 88.0

Table 7: Ablating feature matching on Aachen Day-Night v1.1 using the top-20 reference images retrieved with EigenPlaces [ 8 ] and the E5+1 solver. We report recalls (higher is better) at pose error thresholds of (0.25 m, 2°) / (0.5 m, 5°) / (5 m, 10°).

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 5 — Figure 5: Gaussian blur and pixelization examples. From left to right: The original image, blur with a Gaussian kernel of size 81px, and a pixelization with 20x downsampling.

Panel of Figure 6 — Figure 6: Selective anonymization example. From top left to right: The original image, the segmentation of privacy-revealing classes, easy-anon - single and easy-anon - inpaint .

Panel of Figure 7 — Figure 7: Comparison of segmentation masks on the Indoor-6 dataset. Left to right: original image , SAM1 - fine masks segmentation and Mask2Former - semantic segmentation (using ADE20k classes).

Panel of Figure 8 — Figure 8: Examples of the used edge extraction methods. From left to right: Canny edge detector used directly on the image, Canny edge detector applied on a Metric3D depth map, and an edge map extracted with DiffusionEdge.

Panel of Figure 9 — Figure 9: Comparison of Canny and SAM1 - fine borders on Cambridge Landmarks dataset. On the top row: original image and Canny edge map. The bottom row contains two maps generated with SAM1 - fine borders . The right one was generated with text filtration.

		Great Court			King’s College			Old Hospital			Shop Façade			St Mary’s Church
		MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.
method		[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]
E5+1	original images	0.29	0.13	43.2	0.19	0.30	60.6	0.24	0.48	51.6	0.06	0.28	95.1	0.10	0.31	86.8
	blur - 41 px	0.41	0.24	32.6	0.21	0.31	56.3	0.29	0.51	45.1	0.07	0.31	95.1	0.13	0.36	81.1
	blur - 81 px	1.12	0.94	7.4	0.43	0.76	22.2	0.52	1.06	17.0	0.19	0.66	68.0	0.44	1.34	24.5
	pixelization - 10x	0.65	0.39	22.5	0.37	0.63	31.5	0.61	1.11	24.2	0.10	0.42	87.4	0.15	0.50	75.7

Appendix

Table 8: Localization results on Cambridge Landmarks [ 44 ] using the top-20 reference images retrieved with EigenPlaces [ 8 ] , RoMa [ 33 ] feature matching, and the E5+1 pose solver. We report median position (MPE) and orientation (MOE) errors (smaller is better) and recall (rec.) at 25 cm, 2° pose error (higher is better).

	Bedroom			Living Room (1)			Living Room (2)			Office
	Table			Pillows			Sofa			Chairs
	MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.
obfuscation method	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]
original images	0.01	0.42	93.5	0.01	0.37	94.6	0.01	0.22	98.5	0.01	0.26	89.4
blur - 41 px	0.04	0.93	60.0	0.02	0.55	83.3	0.02	0.51	75.4	0.03	0.51	70.0
blur - 81 px	0.10	2.20	19.2	0.05	1.37	53.8	0.06	1.55	45.8	0.09	1.61	26.7

Appendix

Table 9: Localization results on the Remove 360 [ 48 , 49 ] dataset, using the top-20 reference images retrieved with EigenPlaces [ 8 ] , RoMa [ 33 ] feature matching, and the E5+1 pose solver. We reporting median position (MPE) and orientation (MOE) errors (smaller is better) and recall (rec.) at 5 cm, 5° pose error (higher is better).

obfuscation		centroid	day	night
SAM1	fine masks	coord. avg.	63.0 / 80.6 / 94.1	49.7 / 73.8 / 96.3
	fine masks	quad. center	63.1 / 80.6 / 94.1	48.7 / 73.8 / 95.8
	fine borders	coord. avg.	68.6 / 84.1 / 95.3	54.5 / 82.2 / 95.8
	fine borders	quad. center	68.9 / 84.1 / 95.3	54.5 / 81.7 / 95.8

Appendix

Table 10: Ablation on segment centroid computation method used for pose refinement. Showing results on Aachen Day-Night v1.1 [ 110 , 87 , 89 ] , using the top-20 retrieved reference images with EigenPlaces [ 8 ] and RoMa [ 33 ] feature matches for pose estimation with E5+1 and segment centroid matches for pose refinement. We reporting localization recalls (higher is better) at the pose error thresholds of (0.25 m, 2°) / (0.5 m, 5°) / (5 m, 10°).

obfuscation		refined	day	night
original images		✗	77.9 / 90.5 / 98.2	64.9 / 88.5 / 98.4
SAM2	coarse masks	✗	28.6 / 46.8 / 77.7	17.8 / 34.0 / 70.2
	coarse masks	✓	31.3 / 47.3 / 75.0	18.3 / 36.6 / 67.0
	coarse borders	✗	36.2 / 50.1 / 78.2	20.9 / 36.1 / 70.7
	coarse borders	✓	35.3 / 51.1 / 76.9	21.5 / 37.7 / 67.5
	fine masks	✗	52.1 / 69.8 / 90.5	30.9 / 56.0 / 85.9

Appendix

Table 11: Localization results on Aachen Day-Night v1.1 [ 110 , 87 , 89 ] using the top-20 retrieved reference images with EigenPlaces [ 8 ] and RoMa [ 33 ] feature matches for pose estimation with E5+1 and segment centroid matches for pose refinement. We reporting localization recalls (higher is better) at the pose error thresholds of (0.25 m, 2°) / (0.5 m, 5°) / (5 m, 10°).

		scene1			scene2a			scene3			scene4a			scene5			scene6
		MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.
method		[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]
E5+1	original images	0.01	0.16	92.9	0.01	0.11	94.2	0.01	0.12	97.1	0.01	0.17	96.8	0.01	0.19	92.7	0.01	0.12	95.7
	blur - 41 px	0.02	0.27	84.7	0.02	0.19	85.2	0.01	0.21	92.1	0.01	0.31	91.8	0.03	0.42	75.0	0.01	0.19	93.5
	blur - 81 px	0.04	0.70	63.1	0.05	0.53	51.4	0.03	0.63	69.8	0.03	0.81	63.3	0.07	1.08	36.1	0.02	0.42	82.4
	pixelization - 10x	0.02	0.31	87.4	0.02	0.26	76.3	0.01	0.27	90.5	0.02	0.41	84.8	0.03	0.50	73.3	0.01	0.23	91.0

Appendix

Table 12: Localization results on Indoor-6 [ 25 ] using the top-20 reference images retrieved with EigenPlaces [ 8 ] , RoMa [ 33 ] feature matching, and the E5+1 pose solver as well as local triangulation. We report median position (MPE) and orientation (MOE) errors (smaller is better) and recall (rec.) at 5 cm, 5° pose error (higher is better).

			scene1			scene2a			scene3			scene4a			scene5			scene6			average
		ref. poses	MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.
method		ref. poses	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]
E5+1	easy-anon - single	obf.	0.03	0.51	73.1	0.03	0.36	66.5	0.04	0.70	59.7	0.02	0.68	79.1	0.04	0.61	61.6	0.02	0.39	84.8	0.03	0.54	70.80
E5+1	SAM1 - fine borders	obf.	0.05	1.14	49.7	0.07	0.56	31.1	0.04	0.68	63.2	0.03	0.77	60.8	0.05	0.87	48.6	0.04	0.94	56	0.05	0.83	51.57
LT	easy-anon - single	obf.	0.04	0.56	64.2	0.04	0.35	65.4	0.05	0.82	55.2	0.05	1.14	51.3	0.05	0.88	48.8	0.02	0.45	78.3	0.04	0.70	60.5
LT	SAM1 - fine borders	obf.	0.06	1.34	38.3	0.07	0.58	30.0	0.05	1.02	50.2	0.06	1.37	43.7	0.07	1.16	36.6	0.05	1.10	44.9	0.06	1.09	40.6

Appendix

Table 13: Localization results on Indoor-6 [ 25 ] using "end-to-end" pipeline with reference poses from SfM on obfuscated images (obf.), using top-10 reference images retrieved with EigenPlaces [ 8 ] , SuperPoint featuress with LightGlue matching, and pose estimation with E5+1 and local triangulation (LT). Reporting median position (MPE) and orientation (MOE) errors (smaller is better) and recall (rec.) at 5 cm, 5° pose error (higher is better).

Panel of Figure 10 — Figure 10: Reconstructions of the St Mary’s Church scene from Cambridge Landmarks [ 44 ] from original images using SuperPoint [ 23 ] features (left) and ALIKED [ 111 , 112 ] features (right). We can see that in the right image, one side of the cathedral collapsed onto the other.

method	retr.	day	night
original images	orig.	82.4 / 93.9 / 99.2	70.7 / 89.0 / 98.4
easy-anon	orig.	82.8 / 93.8 / 99.2	72.8 / 87.4 / 98.4
- single	obf.	82.5 / 93.7 / 99.0	71.7 / 89.0 / 98.4
easy-anon	orig.	82.4 / 93.8 / 99.0	73.3 / 88.0 / 98.4
- inpaint	obf.	81.7 / 93.8 / 99.0	71.2 / 89.5 / 98.4
SAM2	orig.	52.1 / 69.8 / 90.5	30.9 / 56.0 / 85.9

Appendix

Table 14: Localization results on Aachen Day-Night v1.1 [ 110 , 87 , 89 ] using the top-20 reference images retrieved with EigenPlaces [ 8 ] , RoMa [ 33 ] feature matching, and pose estimation with the E5+1 solver. The retrieval is done either on the original images (orig.) or on the obfuscated images (obf.). We report the percentage of queries localized within error thresholds of (0.25 m, 2°) / (0.5 m, 5°) / (5 m, 10°).

matcher	day	night
original images
RoMa	77.9 / 90.5 / 98.2	64.9 / 88.5 / 98.4
RoMa v2	83.0 / 94.9 / 99.5	71.2 / 88.0 / 99.0
Tiny RoMa	77.3 / 88.6 / 97.1	56.0 / 74.9 / 97.4
XFeat	55.7 / 75.6 / 94.3	18.8 / 42.9 / 87.4
Disk + LG	80.0 / 92.6 / 98.8	68.1 / 85.9 / 97.9

Appendix

Table 15: Comparing results of pose estimation with E5+1 and different feature matching methods on Aachen Day-Night v1.1 [ 110 , 87 , 89 ] using top-20 retrieved reference images with EigenPlaces [ 8 ] . Reporting localization recalls (higher is better) at the pose error thresholds of (0.25 m, 2°) / (0.5 m, 5°) / (5 m, 10°).

obfuscation	ms per image
pixelization - 20x	19
pixelization - 10x	20
Canny	32
blur - 81 px	56
blur - 41 px	133
easy-anon - single	255

Appendix

Table 16: Runtime comparison of different obfuscation methods averaged on a random subset of 500 images from the Aachen Day-Night v1.1 [ 110 , 87 , 89 ] dataset. All measurements were performed on a machine with NVIDIA A40 GPU and AMD EPYC 7543 CPU. SAM with filtered text is denoted as (nt) - no text.

					Bedroom			Living Room (1)			Living Room (2)			Office
					Table			Pillows			Sofa			Chairs
					MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.	MPE	MOE	rec.
query		ref.		matching	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]	[m]	[°]	[%]
images	orig.	images	orig.	RoMa	0.01	0.42	93.5	0.01	0.37	94.6	0.01	0.23	98.5	0.02	0.26	89.4
				SP+LG	0.02	0.40	91.2	0.01	0.37	94.1	0.01	0.27	92	0.02	0.32	82.1
				ALIKED+LG	0.02	0.44	86.2	0.01	0.40	90	0.01	0.27	89.4	0.02	0.36	73

Appendix

Table 17: Local feature matching ablation on the Remove 360 [ 48 , 49 ] dataset, using the top-20 reference images retrieved with EigenPlaces [ 8 ] , and the E5+1 pose solver. We reporting median position (MPE) and orientation (MOE) errors (smaller is better) and recall (rec.) at 5 cm, 5° pose error (higher is better).

	obfuscation	runtime
E5+1	original images	24:29
E5+1	SAM1 - fine borders	1:08:51
LT	original images	1:36:00
LT	SAM1 - fine borders	3:41:57

Appendix

Table 18: Runtimes for the E5+1 and local triangulation (LT) pipeline on the Aachen Day-Night v1.1 [ 110 , 87 , 89 ] dataset using top-20 retrieved reference images with EigenPlaces [ 8 ] . Experiments were performed on a machine with NVIDIA A40 GPU and AMD EPYC 7543 CPU.

type		value / unit
original images	orig.	380 kB / image
	PNG (q=75)	1058 kB / image
	JPG (q=80)	392 kB / image
SAM1 - fine masks	PNG (q=75)	366 kB / image
	JPG (q=80)	416 kB / image
SAM1 - fine borders	PNG (q=75)	257 kB / image

Appendix

Table 19: Data size for different image and feature types. The data for images are averages over the Aachen Day-Night v1.1 query set at the original image size. The first row marked with "orig." are the original dataset images, where we do not have control over the used format and encoding. The "(q=N)" values specify the "-quality" parameter used during conversion with ImageMagick [ 43 ] . The values for features are theoretical data sizes based on the dimensionality of their descriptors (128-D for ALIKED and 256-D for SuperPoint) and the used data type (16-bit float). Note that the local features also contain an additional 4 bytes for keypoint coordinates.

Camera Pose Estimation

Camera pose estimation recovers a camera’s position and orientation from images, visual features, scene geometry, and sometimes sensor measurements.

See all 241 papers

Revisiting Geometric Obfuscation with Dual Convergent Lines for Privacy-Preserving Image Queries in Visual Localization
Apr 24, 2026Jeonggon Kim, Heejoon Moon, Je Hyeong HongCamera Pose Estimation Privacy-Preserving ML
Dept. Electronic Engineering, Hanyang University · Dept. Artificial Intelligence, Hanyang University
Depth-Guided Privacy-Preserving Visual Localization Using 3D Sphere Clouds
May 1, 2026Heejoon Moon, Jongwoo Lee, Jeonggon Kim +1Camera Pose Estimation 3D Scene Representation
Department of Artificial Intelligence Hanyang Univeristy Seoul, Republic of Korea · Department of Electronic Engineering Hanyang Univeristy Seoul, Republic of Korea
Bit-ViP: Leveraging Bit-planes to Preserve Visual Privacy in Images through Obfuscation
Jun 28, 2026Vishesh Kumar Tanwar, Ashish Gupta, Sanjay Madria +1Privacy-Preserving ML
Department of Computer Science, Missouri University of Science and Technology, USA · Department of Computer Science and Engineering, BITS Pilani Dubai Campus, Dubai, UAE

Apr 24, 2026J/K move · Enter open · S save

Revisiting Geometric Obfuscation with Dual Convergent Lines for Privacy-Preserving Image Queries in Visual Localization

Jeonggon Kim, Heejoon Moon, Je Hyeong Hong

Dept. Electronic Engineering, Hanyang University · Dept. Artificial Intelligence, Hanyang University

Camera Pose Estimation Privacy-Preserving ML

Open paper AskPDF arXiv

Privacy-Preserving Image Queries (PPIQ) are an emerging mechanism for cloud-based visual localization, enabling pose estimation from obfuscated features instead of private images or raw keypoints. However, the main approaches for PPIQ, primarily geometry-based and segmentation-based obfuscation, both suffer from vulnerabilities to recent privacy attacks. In particular, a fundamental limitation of geometry-based obfuscation is that the spatial distribution of obfuscated neighboring lines still effectively surrounds the original keypoint location, providing exploitable cues for recovering the original points. We revisit this geometric paradigm and introduce Dual Convergent Lines (DCL), a novel keypoint obfuscation method demonstrating strong resilience against such attack. DCL places two fixed anchors on a central partition line and lifts each keypoint to a line originating from one of them, with the active anchor determined by the keypoint's location. This arrangement invalidates the geometry-recovery attack by making its optimization ill-posed: Neighboring lines either misleadingly converge to one anchor, yielding a trivial solution, or become near-parallel at the partition boundary, yielding an unstable high-variance solution. Both outcomes thwart point recovery. DCL is also compatible with an existing line-based solver, enabling deployment in traditional localization pipelines. Experiments on both indoor and large-scale outdoor datasets demonstrate DCL's robustness against privacy attacks, efficiency, and scalability, while achieving practical localization performance.

Abstract

Camera Pose Estimation

Explore similar work

Revisiting Geometric Obfuscation with Dual Convergent Lines for Privacy-Preserving Image Queries in Visual Localization

Depth-Guided Privacy-Preserving Visual Localization Using 3D Sphere Clouds

Bit-ViP: Leveraging Bit-planes to Preserve Visual Privacy in Images through Obfuscation