Event-Aided Sharp Radiance Field Reconstruction for Fast-Flying Drones

Rong Zou; Marco Cannici; Davide Scaramuzza

cs.CVFeb 24, 2026

Event-Aided Sharp Radiance Field Reconstruction for Fast-Flying Drones

Authors: Rong Zou, Marco Cannici, Davide Scaramuzza

Organizations: Robotics and Perception Group, University of Zurich, Switzerland

Abstract

Fast-flying aerial robots promise rapid inspection under limited battery constraints, with direct applications in infrastructure inspection, terrain exploration, and search and rescue. However, high speeds lead to severe motion blur in images and induce significant drift and noise in pose estimates, making dense 3D reconstruction with Neural Radiance Fields (NeRFs) particularly challenging due to their high sensitivity to such degradations. In this work, we present a unified framework that leverages asynchronous event streams alongside motion-blurred frames to reconstruct high-fidelity radiance fields from agile drone flights. By embedding event-image fusion into NeRF optimization and jointly refining event-based visual-inertial odometry priors using both event and frame modalities, our method recovers sharp radiance fields and accurate camera trajectories without ground-truth supervision. We validate our approach on both synthetic data and real-world sequences captured by a fast-flying drone. Despite highly dynamic drone flights, where RGB frames are severely degraded by motion blur and pose priors become unreliable, our method reconstructs high-fidelity radiance fields and preserves fine scene details, delivering a performance gain of over 50% on real-world data compared to state-of-the-art methods.

Figures & tables

Fig. 2 : Overview of our proposed architecture. We reconstruct a sharp radiance field from a set of motion-blurred RGB frames and events. The estimated trajectory from an event-based SLAM system, $T_{init}$ , is refined through a learned Pose Refiner that takes a query timestamp $t$ as input and predicts residual corrections at arbitrary time resolutions. This refined trajectory is used to supervise the radiance field through three complementary branches: a blur branch models image formation across the exposure time; an event branch supervises high-temporal-resolution brightness changes via a learned event camera response function; and a prior branch introduces model-based deblur constraints. All branches share the same learned camera trajectory, allowing events and images to jointly refine the scene representation and motion estimates.

Fig. 3 : Data collection setup. A beamsplitter with an RGB and an event camera is attached to a quadrotor platform.

Method	Factory			Pool			Tanabata			Trolley			Average
Method	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$
DeblurNeRF [ 25 ]	24.52	0.25	0.79	26.02	0.34	0.69	21.38	0.28	0.71	23.58	0.22	0.79	23.87	0.27	0.74
BAD-NeRF [ 28 ]	21.20	0.22	0.64	27.13	0.23	0.70	20.89	0.25	0.65	22.76	0.18	0.73	22.99	0.22	0.68
PDRF [ 26 ]	27.34	0.17	0.87	27.46	0.32	0.72	24.27	0.20	0.81	26.09	0.15	0.86	26.29	0.21	0.81
DP-NeRF [ 27 ]	26.77	0.20	0.85	29.58	0.24	0.79	27.32	0.11	0.85	27.04	0.14	0.87	27.68	0.17	0.84
MPRNet [ 44 ] + NeRF	19.09	0.37	0.56	25.49	0.39	0.64	17.79	0.42	0.51	19.82	0.31	0.62	20.55	0.37	0.58

TABLE I : Quantitative comparison on the synthetic Ev-DeblurBlender dataset. Best results are reported in bold.

Method	Batteries			Power supplies			Lab Equipment			Drones			Figures			Average
Method	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$
DP-NeRF+TensoRF [ 17 ]	26.64	0.27	0.81	25.74	0.32	0.77	27.49	0.31	0.80	26.52	0.30	0.81	27.76	0.34	0.77	26.83	0.31	0.79
EDI [ 8 ] + NeRF	28.66	0.12	0.87	28.16	0.09	0.88	31.45	0.13	0.89	29.37	0.10	0.88	31.44	0.12	0.88	29.82	0.11	0.88
E 2 NeRF [ 43 ]	30.57	0.12	0.88	29.98	0.11	0.87	30.41	0.16	0.86	30.41	0.14	0.87	31.03	0.14	0.85	30.48	0.13	0.87
Ev-DeblurNeRF [ 12 ]	33.17	0.05	0.92	32.35	0.06	0.91	33.01	0.08	0.91	32.89	0.05	0.92	33.39	0.07	0.90	32.96	0.06	0.91
Ours	33.20	0.08	0.92	32.49	0.07	0.91	32.83	0.11	0.91	33.00	0.07	0.92	33.63	0.11	0.90	33.03	0.09	0.91

TABLE II : Quantitative comparison on the real-world Ev-DeblurCDAVIS dataset. Best results are reported in bold.

Panel of Figure 5 — Fig. 4 : Factory (top) and Trolley (bottom) trajectory analysis on NoisyPose-EvDeblurBlender at noise level 4. We present the change of trajectory errors with traveled distance (left), for both translation (meters and percentage) and rotation (degrees and degrees per meter), along with a visual comparison of the trajectories (right).

Panel of Figure 6 — Fig. 5 : Novel view synthesis comparison on the Ev-DeblurCDAVIS dataset.

Exposure	Poses	Method	Max Speed 0.8 $\sim$ 1 m/s			Max Speed 1.2 $\sim$ 1.4 m/s			Max Speed 1.4 $\sim$ 1.8 m/s			Average
Exposure	Poses	Method	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$
10ms	USLAM	Ev-DeblurNeRF [ 12 ]	16.92	0.57	0.45	16.30	0.59	0.42	15.99	0.64	0.42	16.40	0.60	0.43
	USLAM	Ours	28.14	0.21	0.77	27.34	0.23	0.76	26.69	0.27	0.74	27.39	0.24	0.76
	our refined	Ev-DeblurNeRF [ 12 ]	24.03	0.24	0.72	24.28	0.26	0.72	23.17	0.32	0.69	23.83	0.27	0.71
	our refined	E 2 NeRF [ 43 ]	23.77	0.24	0.71	23.47	0.30	0.68	21.46	0.37	0.62	22.90	0.30	0.67
30ms	USLAM	Ev-DeblurNeRF [ 12 ]	16.49	0.46	0.43	16.33	0.51	0.45	15.69	0.55	0.43	16.17	0.51	0.44

TABLE V : Quantitative comparison on the new Gen3-HandHeld dataset, and ablation of speed vs. exposure time. USLAM stands for UltimateSLAM [ 33 ] .

Panel of Figure 8 — Fig. 6 : Novel view synthesis comparison on the NoisyPose-EvDeblurBlender dataset under noise level four. We indicate frame-only methods with (F) and frames+events with (F+E), while Train view refers to the closest blurry image in the training set.

Panel of Figure 9 — Fig. 7 : Novel view synthesis comparison on the Gen3-HandHeld dataset under the fastest speed profile. We indicate in parentheses the poses each method uses as initialization, $T_{\text{ref}}$ for our refined poses, $T_{\text{USLAM}}$ for UltimateSLAM [ 33 ] poses.

Panel of Figure 10 — Fig. 8 : Novel view synthesis comparison on the Gen3-DroneFlight dataset under the fastest speed profile. We indicate in parentheses the poses each method uses as initialization, $T_{\text{ref}}$ for our refined poses, and $T_{\text{USLAM}}$ for UltimateSLAM [ 33 ] poses.

Sequence	Poses	Method	Max Speed 1 m/s			Max Speed 1.5 m/s			Max Speed 2 m/s			Average
Sequence	Poses	Method	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$
BoxStack	USLAM	Ev-DeblurNeRF [ 12 ]	13.34	0.58	0.30	12.08	0.64	0.25	13.58	0.66	0.30	13.00	0.63	0.28
	USLAM	Ours	23.89	0.21	0.68	23.49	0.26	0.67	23.05	0.33	0.66	23.48	0.26	0.67
	our refined	Ev-DeblurNeRF [ 12 ]	21.83	0.23	0.61	22.37	0.27	0.63	21.12	0.36	0.59	21.78	0.29	0.61
	our refined	E 2 NeRF [ 43 ]	22.23	0.28	0.61	18.41	0.45	0.45	14.38	0.62	0.33	18.34	0.45	0.46
	our refined	Ours	24.40	0.21	0.71	23.58	0.25	0.67	23.31	0.30	0.67	23.76	0.26	0.68

TABLE VI : Quantitative comparison on the new Gen3-DroneFlight dataset. USLAM stands for UltimateSLAM [ 33 ] .

Poses	Method	Scene 1 (15 m Traj.)			Scene 2 (15 m Traj.)			Scene 1+2 (30 m Traj.)			Average
Poses	Method	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$
USLAM	Ev-DeblurNeRF [ 12 ]	11.48	0.78	0.18	11.77	0.70	0.13	10.22	0.77	0.15	11.16	0.75	0.15
USLAM	Ours	24.81	0.24	0.70	25.70	0.29	0.72	25.59	0.25	0.74	25.37	0.26	0.72
our refined	Ev-DeblurNeRF [ 12 ]	24.09	0.28	0.68	26.46	0.25	0.76	25.93	0.24	0.75	25.49	0.26	0.73
our refined	E 2 NeRF [ 43 ]	21.81	0.32	0.62	24.40	0.31	0.68	24.17	0.27	0.70	23.46	0.30	0.67
our refined	Ours	25.11	0.22	0.72	26.62	0.24	0.75	26.13	0.23	0.75	25.95	0.23	0.74

TABLE VII : Quantitative comparison on trajectories of different lengths. All trajectories were flown by a real drone under the fastest speed profile. Scene 1: Models , Scene 2: ReadingCorner .

$\mathcal{L}_{\text{blur}}$	$\mathcal{L}_{\text{prior}}$	$\mathcal{L}_{\text{event}}$	Max Speed 1 m/s			Max Speed 1.5 m/s			Max Speed 2 m/s			Average
$\mathcal{L}_{\text{blur}}$	$\mathcal{L}_{\text{prior}}$	$\mathcal{L}_{\text{event}}$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$
✓			23.60	0.44	0.66	18.65	0.60	0.46	15.74	0.68	0.36	19.33	0.57	0.49
✓		✓	25.10	0.37	0.71	21.75	0.52	0.59	23.01	0.46	0.63	23.29	0.45	0.64
	✓		23.81	0.43	0.67	20.53	0.57	0.55	21.25	0.55	0.56	21.86	0.52	0.59
	✓	✓	24.73	0.35	0.74	23.53	0.41	0.71	21.75	0.47	0.62	23.34	0.41	0.69
✓	✓		24.11	0.42	0.67	21.48	0.52	0.57	21.86	0.51	0.58	22.49	0.48	0.61

TABLE VIII : Ablation study on the technical components of our method.

Neural Radiance Fields

Neural radiance fields represent 3D scenes as continuous functions of position and viewing direction to synthesize images from new camera views.

See all 82 papers

Tiny-DroNeRF: Tiny Neural Radiance Fields aboard Federated Learning-enabled Nano-drones
Mar 2, 2026Ilenia Carboni, Elia Cereda, Lorenzo Lamberti +3Neural Radiance Fields 3D Reconstruction
Dalle Molle Institute for Artificial Intelligence, USI-SUPSI, Switzerland. · Department of Electrical, Electronic and Information Engineering, University of Bologna, Italy. · Integrated Systems Laboratory, ETH Zürich, Switzerland.
AsyncEvGS: Asynchronous Event-Assisted Gaussian Splatting for Handheld Motion-Blurred Scenes
May 8, 2026Jun Dai, Renbiao Jin, Bo Xu +5Event-Based Vision 3D Gaussian Splatting
Shanghai AI Laboratory · Shanghai Jiao Tong University · CUHK MMLab +1
On-the-Fly3R: Towards Robust Online 3D Reconstruction with Feed-Forward 3R Models for Large-Scale UAV Scenarios
Sep 1, 2026Zhe Shen, Liyuan Lou, Yifei Yu +43D Reconstruction Feed-Forward 3D Reconstruction
School of Geodesy and Geomatics, Wuhan University, Wuhan 430079, China · School of Geodesy

Mar 2, 2026J/K move · Enter open · S save

Tiny-DroNeRF: Tiny Neural Radiance Fields aboard Federated Learning-enabled Nano-drones

Ilenia Carboni, Elia Cereda, Lorenzo Lamberti, Daniele Malpetti+2

Dalle Molle Institute for Artificial Intelligence, USI-SUPSI, Switzerland. · Department of Electrical, Electronic and Information Engineering, University of Bologna, Italy. · Integrated Systems Laboratory, ETH Zürich, Switzerland.

Neural Radiance Fields 3D Reconstruction Tiny ML Aerial Robotics

Open paper AskPDF arXiv

Sub-30g nano-sized aerial robots can leverage their agility and form factor to autonomously explore cluttered and narrow environments, like in industrial inspection and search and rescue missions. However, the price for their tiny size is a strong limit in their resources, i.e., sub-100 mW microcontroller units (MCUs) delivering

\sim

100 GOps/s at best, and memory budgets well below 100 MB. Despite these strict constraints, we aim to enable complex vision-based tasks aboard nano-drones, such as dense 3D scene reconstruction: a key robotic task underlying fundamental capabilities like spatial awareness and motion planning. Top-performing 3D reconstruction methods leverage neural radiance fields (NeRF) models, which require GBs of memory and massive computation, usually delivered by high-end GPUs consuming 100s of Watts. Our work introduces Tiny-DroNeRF, a lightweight NeRF model, based on Instant-NGP, and optimized for running on a GAP9 ultra-low-power (ULP) MCU aboard our nano-drones. Then, we further empower our Tiny-DroNeRF by leveraging a collaborative federated learning scheme, which distributes the model training among multiple nano-drones. Our experimental results show a 96% reduction in Tiny-DroNeRF's memory footprint compared to Instant-NGP, with only a 5.7 dB drop in reconstruction accuracy. Finally, our federated learning scheme allows Tiny-DroNeRF to train with an amount of data otherwise impossible to keep in a single drone's memory, increasing the overall reconstruction accuracy. Ultimately, our work combines, for the first time, NeRF training on an ULP MCU with federated learning on nano-drones.

Abstract

Neural Radiance Fields

Explore similar work

Tiny-DroNeRF: Tiny Neural Radiance Fields aboard Federated Learning-enabled Nano-drones

AsyncEvGS: Asynchronous Event-Assisted Gaussian Splatting for Handheld Motion-Blurred Scenes

On-the-Fly3R: Towards Robust Online 3D Reconstruction with Feed-Forward 3R Models for Large-Scale UAV Scenarios