VkVIO: Cross-platform GPU Acceleration for Visual-Inertial Odometry with Vulkan

Ole Hoffmann; Mateo de Mayo; Daniel Cremers

cs.ROSep 24, 2026

VkVIO: Cross-platform GPU Acceleration for Visual-Inertial Odometry with Vulkan

Authors: Ole Hoffmann, Mateo de Mayo, Daniel Cremers

Organizations: Technical University of Munich, Munich, Germany · Munich Center for Machine Learning, Munich, Germany

Abstract

Perception in robotics and XR fundamentally relies on good state estimation. Visual-inertial odometry (VIO) and Simultaneous Localization and Mapping (VI-SLAM) are proven ways of achieving this goal in a cost-effective and accurate manner. Efficiency in these systems allows for smaller, cooler, and lighter devices. GPU acceleration is a natural approach for reducing latency, thanks to their wide availability in platforms like embedded computers, mobile phones, and XR headsets. However, previous works in the literature have limited themselves to the use of CUDA for this task, significantly reducing deployment options to a single vendor. We instead leverage the vendor-agnostic Vulkan API, originally designed for the strict performance requirements of 3D graphics applications. In this work, we present VkVIO, the first, to the best of our knowledge, cross-platform GPU-accelerated VIO method. We provide state-of-the-art accuracy with causal estimates required for real-time operation. We deploy VkVIO on a diverse range of devices spanning a workstation, a laptop, and an extremely inexpensive single-board computer, while outperforming CUDA-based systems on the same hardware. VkVIO enables possibilities for low-latency, low-power, and low-cost VIO in robotics and XR.

Figures & tables

Fig. 1: VkVIO runs the frontend of a stereo visual-inertial odometry (VIO) pipeline on any Vulkan-capable GPU. Top left: progress on EuRoC MH02 after the same wall-clock time for VkVIO (Vk), Basalt (Bt), Jetson-SLAM (JS), and FastTrack (FT), all executed on the same Nvidia RTX 3070 machine. Top right: end-to-end speedup of VkVIO over Basalt on Nvidia RTX 3070, Apple M3, and Radxa Cubie A7Z for six evaluation sequences, ordered by the pixels each delivers per frame (all cameras, in megapixels). Bottom: general pipeline overview of VkVIO.

Fig. 2: Data flow of the VkVIO frontend for one frame. The CPU uploads the stereo images, the previous keypoints $\mathrm{KP}_{t-1}$ , and the mask buffer. The four stages run as compute shaders from a single command buffer. The outputs $\mathrm{TR}_{t}$ (tracked and cross-matched keypoint positions) and $\mathrm{DET}_{t}$ (newly detected keypoints) are downloaded back to the CPU.

Fig. 3: GPU feature detection, from image to invocation. Each

24\times 24

grid cell (blue) maps to one workgroup, its pixels to subgroups of 32 invocations (pink), and each pixel (black) to one invocation, which runs the four-point early-out, the FAST-9 segment test, and

3\times 3

NMS. A bitonic sort per workgroup keeps the

k

strongest survivors, appended to

\mathrm{DET}_{t}

Fig. 4: GPU feature tracking, one workgroup per keypoint. Forward pass: the reference patch (blue) is sampled from the reference image. The backend’s initial estimate (pink) places it in the target image, where it is iterated until it aligns with the reference patch (blue). Backward pass: the same steps with the images exchanged (green). The returned position is validated against the original and written to the output buffer $\mathrm{TR}_{t}$ . Dashed connectors indicate which image each block reads.

Fig. 5: Input imagery of the six capture devices in the evaluation. Each panel shows one frame from every camera of a device, labeled with camera count, resolution, and frame rate, and all panels are drawn to a common pixel scale so that their area reflects the pixels delivered per frame.

		ATE (m)		RTE (m)		Time (ms)
Seq.	Sys.	mean	med.	mean	med.	FE	E2E
MOO*	Bt	0.213	0.063	0.009	0.008	$3.6\pm 0.2$	$4.8\pm 0.3$
16 seq.	Vk	0.160	0.064	0.009	0.008	$1.9\pm 0.1$	$3.1\pm 0.3$
MGO*	Bt	0.220	0.075	0.014	0.013	$7.0\pm 0.4$	$8.6\pm 0.3$
15 seq.	Vk	0.221	0.080	0.014	0.013	$3.0\pm 0.2$	$4.6\pm 0.5$
MIO*	Bt	0.312	0.088	0.014	0.007	$8.4\pm 0.4$	$9.7\pm 0.5$

TABLE I: Accuracy maintained from Basalt baseline

Fig. 6: Per-stage frontend time of Basalt (Bt) and VkVIO (Vk) on the evaluation sequences, ordered from fewer to more input pixels. Stages include image pyramid, feature tracking, feature detection, and feature matching.

Fig. 7: Per-frame frontend time of Basalt and VkVIO on HF151, illustrating the more stable timing of the GPU frontend. Stacked areas show the 50-frame rolling mean of each stage.

FPS

E2E

p99

Platf.

Seq.

Spd.

Apple M3

TR2

283

351

2.26\pm 0.16

1.63\pm 0.17

1.4

\times

1.27\pm 0.44

1.22\pm 0.46

3.53\pm 0.46

2.85\pm 0.51

1.2

\times

5.26

4.59

MOO02

209

326

3.61\pm 0.37

1.94\pm 0.06

1.9

\times

1.17\pm 0.72

1.13\pm 0.72

4.77\pm 0.71

3.06\pm 0.71

1.6

\times

6.76

5.23

EMH02

130

204

4.76\pm 1.32

1.93\pm 0.23

2.5

\times

2.95\pm 1.60

2.96\pm 1.47

7.71\pm 1.67

4.89\pm 1.55

1.6

\times

13.18

9.43

MGO02

114

237

7.22\pm 0.51

2.79\pm 0.67

2.6

\times

1.54\pm 0.85

1.44\pm 0.83

8.76\pm 0.90

4.23\pm 1.08

2.1

\times

10.91

6.95

MIO02

101

300

8.58\pm 0.47

2.11\pm 0.51

4.1

\times

1.34\pm 0.76

1.23\pm 0.72

9.92\pm 0.78

3.34\pm 0.87

3.0

\times

11.90

5.53

TABLE II: Cross-Platform Runtime Comparison

Fig. 8: Long-term operation on the MIPB08 endurance sequence on the M3. We show power, temperature, and clock frequency for the CPU (Basalt) and GPU (VkVIO) frontends. Raw samples are drawn shaded under a centered 50 s rolling mean.

	FPS				E2E				ATE				RTE
Seq.	Vk	Bt	JS	FT	Vk	Bt	JS	FT	Vk	Bt	JS	FT	Vk	Bt	JS	FT
TR2	535	287	144	177	$1.87\pm 0.37$	$3.49\pm 0.40$	$6.93\pm 2.76$	$5.66\pm 0.89$	0.053	0.059	0.146	0.026	0.007	0.007	0.117	0.011
MOO02	495	193	159	158	$2.02\pm 0.70$	$5.18\pm 0.75$	$6.30\pm 2.79$	$6.30\pm 1.05$	0.235	0.234	0.392	0.348	0.015	0.015	0.214	0.042
EMH02	269	128	136	172	$3.72\pm 1.40$	$7.82\pm 1.47$	$7.33\pm 4.43$	$5.82\pm 1.47$	0.049	0.055	0.051	0.145	0.004	0.004	0.009	0.050
MGO02 $\dagger$	334	112	197	206	$2.99\pm 0.83$	$8.90\pm 0.91$	$5.07\pm 2.08$	$4.86\pm 1.23$	0.503	0.497	0.654	1.922	0.028	0.028	0.246	0.756
MIO02	379	110	98	150	$2.64\pm 0.70$	$9.06\pm 1.00$	$10.20\pm 2.38$	$6.66\pm 1.20$	1.134	1.133	0.197	5.410	0.056	0.058	0.100	0.191

TABLE III: Comparison with CUDA Frontends

Fig. 9: End-to-end, frontend, and backend times of VkVIO against Basalt and the CUDA-based FastTrack and Jetson-SLAM on EMH02. Raw per-frame times are drawn shaded under a centered 50-frame rolling mean.

Seq.	Sys.	FPS	FE	BE	E2E	p99	ATE	RTE
TR2	Bt	18	$34.2\pm 11.3$	$21.5\pm 11.0$	$55.6\pm 17.8$	109.6	0.068	0.008
	Vk	22	$34.5\pm 2.5$	$12.1\pm 4.5$	$46.6\pm 5.6$	64.9	0.068	0.008
	Vk L	35	$18.8\pm 1.1$	$9.8\pm 4.1$	$28.6\pm 4.6$	42.5	0.082	0.008
MOO02	Bt	12	$62.6\pm 20.1$	$23.8\pm 19.9$	$86.4\pm 31.9$	184.6	0.243	0.015
	Vk	17	$50.2\pm 3.1$	$9.2\pm 5.8$	$59.5\pm 6.1$	76.9	0.240	0.015
	Vk L	32	$20.8\pm 1.7$	$10.0\pm 4.6$	$30.8\pm 4.8$	42.7	0.235	0.015

TABLE IV: VkVIO-lite on the Radxa

Fig. 10: CPU and GPU temperatures for Basalt and VkVIO on the Radxa Cubie A7Z over MOO12, run with no heatsink and no fan. Raw samples are drawn shaded under a centered 50 s rolling mean. VkVIO’s traces stop earlier because it finishes the session sooner.

Visual-Inertial Odometry

Visual-inertial odometry estimates camera motion by combining image geometry with inertial measurements for metric navigation.

See all 57 papers

FastTrack: GPU-Accelerated Tracking for Visual SLAM
Sep 13, 2025Kimia Khabiri, Parsa Hosseininejad, Shishir Gopinath +2GPU Acceleration Visual SLAM
Simon Fraser University · University at Buffalo
GLidE-SLAM: GL-Accelerated Indirect-Direct Embedded SLAM
Jul 18, 2026Carlos A. Pinheiro de Sousa, Heiko Hamann, Oliver DeussenSimultaneous Localization and Mapping Visual Odometry
The authors are with the Department of Computer and Information Science, University of Konstanz, 78464 Konstanz, Germany.
Sphere-VIO: Fast and Robust Visual-Inertial Odometry via Unified Spherical Representation for Heterogeneous Multi-Camera Systems
Jun 29, 2026Yueteng Yang, Yusen Xie, Hao Wei +5Kalman Filtering Error-State Kalman Filtering
The Hong Kong University of Science and Technology (Guangzhou), Guangzhou 511453, China · Differential Robotics, Hangzhou 311100, China · Zhejiang University, Hangzhou 310027, China +1

Sep 13, 2025J/K move · Enter open · S save

FastTrack: GPU-Accelerated Tracking for Visual SLAM

Kimia Khabiri, Parsa Hosseininejad, Shishir Gopinath, Karthik Dantu+1

Simon Fraser University · University at Buffalo

GPU Acceleration Visual SLAM Visual-Inertial SLAM

Open paper AskPDF arXiv

The tracking module of a visual-inertial SLAM system processes incoming image frames and IMU data to estimate the position of the frame in relation to the map. It is important for the tracking to complete in a timely manner for each frame to avoid poor localization or tracking loss. We therefore present a new approach which leverages GPU computing power to accelerate time-consuming components of tracking in order to improve its performance. These components include stereo feature matching and local map tracking. We implement our design inside the ORB-SLAM3 tracking process using CUDA. Our evaluation demonstrates an overall improvement in tracking performance of up to 2.8x on a desktop and Jetson Xavier NX board in stereo-inertial mode, using the well-known SLAM datasets EuRoC and TUM-VI.

Abstract

Visual-Inertial Odometry

Explore similar work

FastTrack: GPU-Accelerated Tracking for Visual SLAM

GLidE-SLAM: GL-Accelerated Indirect-Direct Embedded SLAM

Sphere-VIO: Fast and Robust Visual-Inertial Odometry via Unified Spherical Representation for Heterogeneous Multi-Camera Systems