RoEL: Robust Event-based 3D Line Reconstruction

Gwangtak Bae; Jaeho Shin; Seunggu Kang; Junho Kim; Ayoung Kim; Young Min Kim

cs.ROFeb 20, 2026

RoEL: Robust Event-based 3D Line Reconstruction

Authors: Gwangtak Bae, Jaeho Shin, Seunggu Kang, Junho Kim, Ayoung Kim, Young Min Kim

Organizations: Dept. of Electrical and Computer Engineering, Seoul National University · University of Michigan · Dept. of Future Automotive Mobility, Seoul National University · Dept. of Mechanical Engineering, Seoul National University

Abstract

Event cameras in motion tend to detect object boundaries or texture edges, which produce lines of brightness changes, especially in man-made environments. While lines can constitute a robust intermediate representation that is consistently observed, the sparse nature of lines may lead to drastic deterioration with minor estimation errors. Only a few previous works, often accompanied by additional sensors, utilize lines to compensate for the severe domain discrepancies of event sensors along with unpredictable noise characteristics. We propose a method that can stably extract tracks of varying appearances of lines using a clever algorithmic process that observes multiple representations from various time slices of events, compensating for potential adversaries within the event data. We then propose geometric cost functions that can refine the 3D line maps and camera poses, eliminating projective distortions and depth ambiguities. The 3D line maps are highly compact and can be equipped with our proposed cost function, which can be adapted for any observations that can detect and extract line structures or projections of them, including 3D point cloud maps or image observations. We demonstrate that our formulation is powerful enough to exhibit a significant performance boost in event-based mapping and pose refinement across diverse datasets, and can be flexibly applied to multimodal scenarios. Our results confirm that the proposed line-based formulation is a robust and effective approach for the practical deployment of event-based perceptual modules. Project page: https://gwangtak.github.io/roel/

Figures & tables

Fig. 1 : We present RoEL, an event-based 3D line reconstruction pipeline that achieves noise-robust reconstruction by leveraging line correspondences. Our 3D line maps not only provide efficient representations for edge-capturing event cameras, but also enable cross-modal applications such as registration and localization.

Fig. 2 : Method overview. Our 3D line mapping pipeline, RoEL, consists of two stages: correspondence search and 3D line reconstruction. The first stage takes events and camera poses as input. Through line detection, plane fitting, and matching processes that take into account the characteristics of event data, it finds line correspondences and identifies the events supporting each line. In the second stage, our method triangulates 2D lines to generate initial 3D lines. We further optimize 3D lines with multi-view observations using cost functions defined in 3D space based on the Grassmann distance. Finally, our method outputs 3D line segments that effectively represent the scene.

Fig. 3 : Detection-guided space-time plane fitting. Through plane fitting for each line, raw events are distilled into inlier events, and noisy initial lines are refined. In the center box, the black line depicts an initial line which is used for candidate events selection for plane fitting, the red plane indicates the fitted plane, the red line represents the refined line, and the green dots denote all inlier events. Each line is associated with its corresponding events, and the inaccurate 2D lines are refined, as highlighted in the orange circle.

Fig. 4 : Difference from reprojection error. The two columns illustrate the same error-calculation scenario consisting of 3D line estimates and a 2D observation. In this scenario, different 3D line estimates project to the same 2D location, leading to identical reprojection errors. Our Grassmann-based cost evaluates geometric consistency directly in 3D.

Fig. 5 : Two different cost functions defined for 2D line observations and 2D event observations. The geodesic distance on the Grassmann manifold serve as the basis for cost functions. The left figure illustrates the cost function defined between a 2D line and a 3D line. We use the plane-to-line Grassmann distance between the red plane, which is back-projected from the 2D line, and the blue 3D line. The right figure shows the cost function defined between an event and a 3D line. We use the plane-to-line Grassmann distance between the green line back-projected from the event and the blue plane connecting the 3D line and the camera center.

Parameter	Value
Inlier selection threshold $\tau$	2
Inlier selection $N_{e}$	100
Time-axis scaling for plane fitting	20
Candidate distance to line in Equation 2	10
Epipolar score threshold	0.3
Inlier error threshold using Equation 9	1.4

TABLE I : Key parameters used in our pipeline.

Dataset	Sensor	Event	Pose
Replica [ 62 ]	Synthetic	VID2E [ 63 ]	GT
TUM-VIE [ 64 ]	Prophesee Gen4	Real	DEVO
VECtor [ 65 ]	Prophesee Gen3	Real	DEVO
$I^{2}$ -SLAM [ 66 ]	Synthetic	VID2E [ 63 ]	GT

TABLE II : Summary of datasets and experimental settings.

Method	Scene	Acc.↓	Comp.↓	IoU 5↑	IoU 10↑	IoU 20↑	#Rep.	Rep.
EL-SLAM [ 5 ]	avg	248.92	153.47	0.006	0.021	0.063	1169	Line
[0.7pt/2pt]	room0	194.57	123.34	0.006	0.022	0.063	966
	room1	344.48	99.50	0.004	0.014	0.048	1375
	room2	254.05	177.93	0.007	0.026	0.074	1222
	office0	242.14	81.87	0.009	0.028	0.081	1331
	office1	236.25	72.84	0.006	0.023	0.071	1350

TABLE III : Quantitative comparison of reconstruction quality on the Replica [ 62 ] dataset. ↓: lower is better, ↑: higher is better. Bold indicates best. #Rep. indicates the number of reconstructed lines or points.

Panel of Figure 9 — Fig. 6 : Qualitative comparison of reconstruction quality on the Replica [ 62 ] dataset.

Panel of Figure 10 — Fig. 7 : Qualitative comparison of reconstruction quality on the TUM-VIE [ 64 ] dataset.

Method	Scene	Acc.↓	P 5↑	P 10↑	P 20↑	#Rep.	Rep.
EL-SLAM [ 5 ]	avg	221.85	0.021	0.069	0.140	1711	Line
[0.7pt/2pt]	desk-fast	184.96	0.020	0.062	0.121	1953
	sofa-fast	308.12	0.016	0.052	0.107	1654
	desk-normal	109.17	0.028	0.096	0.197	436
	sofa-normal	285.13	0.021	0.067	0.135	2799
LIMAP [ 15 ]	avg	91.21	0.034	0.105	0.233	278	Line

TABLE IV : Quantitative comparison of reconstruction quality on the VECtor [ 65 ] dataset. Bold indicates best. #Rep. indicates the number of lines or points. We report Accuracy and Precision using the ground-truth 3D point map, as we could not obtain a reliable ground-truth 3D edge map due to misalignment between the RGB and depth images.

Panel of Figure 12 — Fig. 8 : Qualitative comparison of reconstruction quality on desk-normal of the VECtor [ 65 ] dataset. The first column shows the input scene. The ground-truth map, in color, is overlaid on the reconstructed map, represented by black points or lines.

Panel of Figure 13 — Fig. 9 : Qualitative comparison of registration on office4 of the Replica [ 62 ] dataset. The background shows a dense RGB map reconstructed with RGB-D SLAM, and the red-colored map represents the reconstruction results after registration.

Method	Scene	R error ( ∘ ) ↓	t error ( $\mathrm{m}$ ) ↓
EL-SLAM [ 5 ]	avg	4.681	0.271
[0.7pt/2pt]	room0	2.419	0.289
	room1	6.520	0.257
	room2	2.745	0.475
	office0	5.901	0.120
	office1	5.102	0.126

TABLE V : Quantitative comparison of registration performance of the Replica [ 62 ] dataset. Bold indicates best.

Method	Scene	R error ( ∘ ) ↓	t error ( $\mathrm{m}$ ) ↓
EL-SLAM [ 5 ]	avg	15.82	0.62
[0.7pt/2pt]	desk-fast	21.28	0.91
	sofa-fast	26.09	0.88
	desk-normal	6.42	0.28
	sofa-normal	9.49	0.41
LIMAP [ 15 ]	avg	9.42	0.29

TABLE VI : Quantitative comparison of registration performance on the VECtor [ 65 ] dataset. Bold indicates best.

Panel of Figure 16 — Fig. 10 : Qualitative comparison of panoramic localization performance on room0 of the Replica [ 62 ] dataset. The background shows a query RGB panoramic image, and the red lines represent the projected 3D line map at the estimated camera pose. The blue boxes indicate the zoom-in regions.

Method	R error ( ∘ ) ↓	t error ( $\mathrm{m}$ ) ↓	Success (0.1 $\mathrm{m}$ /5 ∘ ) ↑	Success (0.3 $\mathrm{m}$ /15 ∘ ) ↑
EL-SLAM [ 5 ]	104.30	2.37	10.5	30.5
LIMAP [ 15 ]	5.81	0.08	37.0	61.0
Ours	5.18	0.04	45.5	81.5

TABLE VII : Quantitative comparison of panoramic localization performance on the Replica [ 62 ] dataset. Bold indicates best.

LIMAP [ 15 ]	Ours Line Grass.	Ours Event Grass.	ATE ↓ (Synthetic)	ATE ↓ (Real-world)
-	-	-	33.25	0.89
✓	-	-	32.05	0.76
-	✓	-	27.95	0.73
-	✓	✓	6.92	0.67

TABLE VIII : Pose refinement performance analysis. We compare LIMAP’s projection-based line cost and our Grassmann-based line and event cost functions.

Panel of Figure 19 — Fig. 11 : Qualitative comparison of pose refinement performance. Black dashed lines indicates the ground-truth camera trajectory, and the noisy and optimized poses are shown in blur lines.

Panel of Figure 20 — Fig. 12 : Comparison between frame-based line mapping with [ 15 ] and our event-based line mapping under high speed motion and underexposure. Sharp image in favorable condition, blurry image in high speed motion, blurry and underexposed image in high speed motion are used for comparison. Noisy events are synthesized to simulate low-light shot noise by injecting noise events into the clean event stream.

MWMR	Plane Fitting	Opt. Line LIMAP [ 15 ]	Opt. Line Grass.	Opt. Event Grass.	Acc. ↓	Comp. ↓	IoU 20 ↑
-	-	-	-	-	65.98	104.93	0.068
✓	-	-	-	-	68.00	51.17	0.053
✓	✓	-	-	-	47.94	44.09	0.089
✓	✓	✓	-	-	46.57	58.36	0.092
✓	✓	-	✓	-	39.81	58.82	0.097
✓	✓	-	✓	✓	28.63	55.20	0.120

TABLE IX : Ablation study. MWMR refers to the multi-window, multi-representation line detection module. Plane fitting denotes the 2D line refinement step. LIMAP [ 15 ] ’s projection-based line cost and our projection-free line and event cost defined on the Grassmann manifold are also analyzed. When MWMR is not used, we report the average performance over all combinations.

Plane Fitting Method	Precision↑	Recall↑	F-score↑
Direct Plane Fitting	0.603	0.228	0.331
Detection-guided Plane Fitting	0.535	0.526	0.530

TABLE X : Quantitative analysis on plane fitting strategies.

Fig. 13 : Qualitative analysis on plane fitting strategies. The fitted planes are sliced to visualize the resulting line segments.

Multiple Prediction	Line Merging	#Line	Precision↑	Recall↑	F-score↑
-	-	31	0.612	0.344	0.434
✓	-	185	0.359	0.758	0.487
✓	✓	55	0.530	0.523	0.526

TABLE XI : Quantitative analysis across stages of multi-window, multi-representation line detection. When Multiple Prediction is not used, we report the average performance over all predictions.

Window Size	F-score↑	Image Rep.	F-score↑
1/10	0.368	[ 9 ] +TOS	0.502
1/5	0.523	[ 10 ] +TOS	0.523
1	0.526	[ 9 ] + [ 10 ]	0.526
5	0.484	[ 9 ] +SITS	0.499
10	0.443	[ 10 ] +SITS	0.496

TABLE XII : Sensitivity analysis of MWMR line detection with respect to temporal window size and image representations: TOS [ 88 ] and SITS [ 89 , 90 ] .

Image Rep.	#Line	Precision↑	Recall↑	F-score↑
TOS	42	0.527	0.380	0.441
SITS	62	0.443	0.412	0.427
MWMR	55	0.530	0.523	0.526

TABLE XIII : Comparison of velocity-invariant representations and MWMR line detection.

Stage	Time (s)
Line detection	0.064
Plane fitting	0.358
Local matching	0.001
Global matching	0.663
Triangulation	0.120
Optimization	0.351

TABLE XIV : Runtime analysis. The reported values indicate the processing time per frame (in seconds).

Fig. 14 : Example of our method applied to a curved-shape chair. The 2D line detection results from two viewpoints and the reconstructed 3D line map are shown.

3D Reconstruction

3D reconstruction recovers a scene’s shape, appearance, and camera motion from images, video, depth, or other sensor measurements.

See all 955 papers

E2EGS: Event-to-Edge Gaussian Splatting for Pose-Free 3D Reconstruction
Mar 16, 2026Yunsoo Kim, Changki Sung, Dasol Hong +13D Reconstruction Novel View Synthesis
Urban Robotics Lab, School of Electrical Engineering, KAIST
4D Reconstruction with Events & RGB
Oct 27, 2026Dulanga Weerakoon, Darshana Rathnayake, Van Minh Do +74D Reconstruction RGB-Event Sensor Fusion
Singapore-MIT Alliance for Research & Technology Centre · Singapore Management University · Massachusetts Institute of Technology
EventGeM: Global-to-Local Feature Matching for Event-Based Visual Place Recognition
Mar 6, 2026Adam D. Hines, Gokul B. Nair, Nicolás Marticorena +2Visual Place Recognition Event-Based Vision
QUT Centre for Robotics, School of Electrical Engineering and Robotics, Queensland University of Technology, Brisbane, QLD 4000, Australia.

Mar 16, 2026J/K move · Enter open · S save

E2EGS: Event-to-Edge Gaussian Splatting for Pose-Free 3D Reconstruction

Yunsoo Kim, Changki Sung, Dasol Hong, Hyun Myung

Urban Robotics Lab, School of Electrical Engineering, KAIST

3D Reconstruction Novel View Synthesis 3D Gaussian Splatting Camera Pose Estimation

Open paper AskPDF arXiv

The emergence of neural radiance fields (NeRF) and 3D Gaussian splatting (3DGS) has advanced novel view synthesis (NVS). These methods, however, require high-quality RGB inputs and accurate corresponding poses, limiting robustness under real-world conditions such as fast camera motion or adverse lighting. Event cameras, which capture brightness changes at each pixel with high temporal resolution and wide dynamic range, enable precise sensing of dynamic scenes and offer a promising solution. However, existing event-based NVS methods either assume known poses or rely on depth estimation models that are bounded by their initial observations, failing to generalize as the camera traverses previously unseen regions. We present E2EGS, a pose-free framework operating solely on event streams. Our key insight is that edge information provides rich structural cues essential for accurate trajectory estimation and high-quality NVS. To extract edges from noisy event streams, we exploit the distinct spatio-temporal characteristics of edges and non-edge regions. The event camera's movement induces consistent events along edges, while non-edge regions produce sparse noise. We leverage this through a patch-based temporal coherence analysis that measures local variance to extract edges while robustly suppressing noise. The extracted edges guide structure-aware Gaussian initialization and enable edge-weighted losses throughout initialization, tracking, and bundle adjustment. Extensive experiments on both synthetic and real datasets demonstrate that E2EGS achieves superior reconstruction quality and trajectory accuracy, establishing a fully pose-free paradigm for event-based 3D reconstruction.

Abstract

3D Reconstruction

Explore similar work

E2EGS: Event-to-Edge Gaussian Splatting for Pose-Free 3D Reconstruction

4D Reconstruction with Events & RGB

EventGeM: Global-to-Local Feature Matching for Event-Based Visual Place Recognition