ForVis: An In-Field Dataset and Benchmark for VIO Using Under-Canopy UAV Flights in Forests

Arman Kiani; Masoud Ataei; Elvis Gyaase; Jeffrey Eiyike; Aaron Weiskittel; Prabuddha Chakraborty; Vikas Dhiman

cs.ROSep 28, 2026

ForVis: An In-Field Dataset and Benchmark for VIO Using Under-Canopy UAV Flights in Forests

Authors: Arman Kiani, Masoud Ataei, Elvis Gyaase, Jeffrey Eiyike, Aaron Weiskittel, Prabuddha Chakraborty, Vikas Dhiman

Organizations: Department of Electrical and Computer Engineering, University of Maine, Orono, ME 04469, USA · School of Forest Resources, University of Maine, Orono, ME 04469, USA

Abstract

Visual-inertial Simultaneous Localization and Mapping (VI-SLAM) for UAVs remains difficult to evaluate in real forest environments, where motion, illumination changes, repetitive vegetation, and vibration can all affect estimation. We present ForVis, an in-field dataset and benchmark for evaluating VI-SLAM during UAV flight in forest environments. The dataset contains twelve flights across open meadow, above-canopy, and under-canopy conditions in each environment. In total, it provides 563.8s of flight over 1096.8m of trajectory, recorded simultaneously with an Intel RealSense D435i and an OAK-D Pro Wide together with inertial and flight-controller data. We benchmark seven open-source VI-SLAM systems over 504 runs. The results show that sensor choice has a larger effect on trajectory error than the spread between algorithms: all seven methods achieve lower median error on the OAK-D Pro than on the D435i. ForVis is intended to support evaluation of speed, accuracy and robustness for VI-SLAM in challenging forest flight.

Figures & tables

Fig. 1: ForVis records visual-inertial data from two sensors, OAK-D Pro and RealSense D435i, on a UAV. (a) The platform is an ARK-RIG5 inspired drone with an OAK-D Pro W beneath a RealSense D435i on a 5-inch quadrotor, with an ARK X20 RTK GNSS receiver above for ground truth. (b) The same three scenes through both sensors in each of the three environments, annotated with the median FAST corner count per frame — the forest interior is the most feature-rich environment in the dataset and the canopy top the poorest, on both sensors. (c) Seven visual-inertial methods on mdw_1 sequence, aligned to the RTK reference (black). Every method has lower performance on D435i than OAK-D Pro.

Sequence	Roll/Pitch	Yaw	Duration	Path	Alt. range	$v_{\max}$	$\lVert\omega\rVert$ p99	Images	Frames	Size	Closure	Ground truth
	[ ∘ ]	[ ∘ ]	[s]	[m]	[m]	[m/s]	[ ∘ /s]		[%]	[GB]	[m]
mdw_1	31/32	151	39.7	74.0	7.0	4.72	93	7 281	94.4	1.8	0.05	RTK fixed, 2.0 cm
mdw_2	52/37	140	34.7	70.4	13.8	7.99	81	6 159	87.3	1.5	0.10	RTK fixed, 2.0 cm
mdw_3	69/70	427	39.1	66.6	5.6	5.49	158	7 086	91.8	1.8	0.10	RTK fixed, 2.0 cm
abv_1	35/41	386	54.5	93.8	10.8	4.69	117	9 837	91.8	2.0	0.02	RTK fixed, 2.0 cm
abv_2	54/57	464	52.6	108.4	14.1	8.64	98	9 585	93.6	2.1	0.05	RTK fixed, 2.0 cm

TABLE I: The twelve benchmark sequences: three environments, with six sequences under canopy. The roll/pitch column is the total attitude change and yaw is cumulative heading change, both over the whole flight. Frames is the percentage of image frames saved relative the requested rate (30 FPS for D435i and 40 FPS for OAK-D), reported for whichever sensors delivered fewer. Closure is the return-to-start closure: the RTK reference’s own residual for the open-sky sequences, and the pilot’s measured pad return under canopy, measured manually ( † ,

\pm 5

cm), where the receiver drifts further than the aircraft moves.

	Meadow		Above canopy		Under canopy
Method	ATE (m)	Path error	ATE (m)	Path error	ATE (m)	Path error	Avg. runtime (s)	RTF wall / data	CPU mean (%)	Peak mem. (MB)	Seqs completed
OAK-D Pro W
orb_slam3	0.58	0.77%	3.72 2	2.33% 2	1.75	2.25 %	65.0	1.45	238	774	11/12
okvis2_x	0.28	0.38 %	1.24	0.99%	1.82	2.06%	65.7	1.43	234	1016	12/12
basalt	1.71	1.63%	1.03	0.82 %	1.96	2.12%	32.0	0.70	725	241	12/12
stella_vslam	1.10	1.52%	2.54	2.10%	1.88	2.15%	178.5	3.89	174	811	12/12

TABLE II: Accuracy and computational cost for the two sensors. ATE is the median over the sequences of each environment, and path error is the ATE expressed as a fraction of the reconstructed path length. The superscript marks 1 , 2 indicate median over fewer sequences than three and “—” indicates that none of the sequences were tracked completely. The under-canopy columns are scored against a meter-grade reference and are not comparable with the other four. Cost columns are computed over the ten sequences measured under a strictly serial condition (one container, eight physical cores, nothing else running); the two sequences added on 4 September were measured with two workers and are excluded from cost, though not from accuracy. ∗ SVO Pro Pro is fed at capture rate, so its real-time factor is fixed by construction rather than measured.

Sensor	Runs	Usable	Process fail	No ATE	Scale div.	Excursion	Failed
OAK-D Pro W	252	244	3	1	3	1	3%
RealSense D435i	252	228	0	0	18	6	10%
Total	504	472	3	1	21	7	6.3%

TABLE III: Failure summary for the two sensors across all benchmark runs. “Process fail” denotes an execution failure; “No ATE” indicates that a trajectory was produced but could not be evaluated; “Scale div.” denotes scale outside the accepted

0.7

–

1.5

range; and “Excursion” denotes a large transient trajectory error.

Specification	RealSense D435i	OAK-D Pro Wide
Stereo IR Resolution	$848\times 480$	$1280\times 800$
Stereo IR Rate	30 Hz	40 Hz
RGB Resolution	$848\times 480$	$640\times 400$
RGB Rate	30 Hz	20 Hz
Gyroscope Rate	200 Hz	400 Hz
Accelerometer Rate	100 Hz	400 Hz

TABLE IV: Sensor specifications for the ForVis dataset.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

OAK-D Pro W
	Meadow			Above canopy			Under canopy
Method	1	2	3	1	2	3	1	2	3	4	5	6
orb_slam3	0.58	0.66	0.27	—	1.69 2	5.74	1.77	2.05	2.50	1.18	1.73	1.11
okvis2_x	0.43	0.25	0.28	0.76	1.24	6.29	1.66	2.31	2.48	1.36	1.93	1.70
basalt	1.80	0.25	1.71	0.43	1.03	3.55	1.71	2.38	2.32	1.47	2.21	1.67
stella_vslam	0.64	1.45	1.10	1.94	2.54	14.31	1.74	2.23	2.50	1.33	2.01	1.76

Appendix

TABLE V: Absolute trajectory error (SE(3), RMSE, metres) for seven methods on twelve sequences and two cameras. Each entry is the mean over three runs; runs whose scale diverged (outside

\times 0.7

–

1.5

) or which threw a transient excursion are excluded, and a superscript gives the number of runs contributing where fewer than three survived. “—” means no run on that cell produced a usable trajectory. The six under-canopy columns are scored against a meter-grade GNSS reference and are not comparable with the six open-sky columns.

Method	Runs	Usable	Process	No ATE	Scale	Excursion	Failed
ORB-SLAM3	72	68	3	1	0	0	6%
OKVIS2-X	72	72	0	0	0	0	0%
Basalt	72	72	0	0	0	0	0%
Stella-VSLAM	72	67	0	0	5	0	7%
SVO Pro	72	63	0	0	2	7	12%
OpenVINS	72	66	0	0	6	0	8%

Appendix

TABLE VI: Failure breakdown by method across both sensors.

Environment	Runs	Usable	Process	No ATE	Scale	Excursion	Failed
Open meadow	126	121	0	0	3	2	4%
Above canopy	126	101	3	1	18	3	20%
Under canopy	252	250	0	0	0	2	1%

Appendix

TABLE VII: Failure breakdown by environment across all methods and sensors.

Sequence	Environment	Runs	Usable	Crash / No ATE	Guard	Failed
mdw_1	Open meadow	42	39	0	3	7%
mdw_2	Open meadow	42	41	0	1	2%
mdw_3	Open meadow	42	41	0	1	2%
abv_1	Above canopy	42	36	3	3	14%
abv_2	Above canopy	42	35	1	6	17%
abv_3	Above canopy	42	30	0	12	29%

Appendix

TABLE VIII: Failure breakdown by sequence across all methods and sensors.

Explore similar work

May 5, 2026cs.RO

Robust Visual SLAM for UAV Navigation in GPS-Denied and Degraded Environments: A Multi-Paradigm Evaluation and Deployment Study

Reliable localization in GPS-denied, visually degraded environments is critical for autonomous UAV opera- tions. This paper presents a systematic comparative evaluation of five V-SLAM systems ORB-SLAM3, DPVO, DROID-SLAM, DUSt3R, and MASt3R spanning classical, deep learning, recurrent, and Vision Transformer (ViT) paradigms. Experiments are conducted on curated sequences from four public benchmarks (TUM RGB-D, EuRoC MAV, UMA-VI, SubT-MRS) and a custom monocular indoor dataset under five controlled degradation conditions (normal, low light, dust haze, motion blur, and combined), with sub-millimeter Vicon ground truth. Results show that ORB-SLAM3 fails critically under severe degradation (62.4% overall TSR; 0% under dense haze), while learning-based methods remain robust: MASt3R achieves the lowest degraded ATE (0.027 m) and DUSt3R the highest tracking success (96.5%). DPVO offers the best efficiency robustness trade-off (18.6 FPS, 3.1 GB GPU memory, 86.1% TSR), making it the preferred choice for memory-constrained embedded platforms. Embedded deployment analysis across NVIDIA Jetson platforms provides actionable guidelines for SLAM selection under SWaP-constrained UAV scenarios.

Prasoon Kumar, Akshay Deepak, Sandeep Kumar

National Institute of Technology Patna, Bihar, India · Central Research Laboratory, Bharat Electronics Ltd., Ghaziabad, Uttar Pradesh, India

Orb-Slam2 Simultaneous Localization And Mapping Unmanned Aerial Vehicles Global Navigation Satellite System Nvidia Jetson Orin Nano Multi-Dimensional Evaluation

Ask PDF

May 8, 2026cs.RO

AERO-VIS: Asynchronous Event-based Real-time Onboard Visual-Inertial SLAM

The robustness of event cameras to high dynamic range and motion blur holds the potential to improve visual odometry systems in challenging environments. Although their high temporal resolution does not require synchronous processing, most event-based odometry methods still run at fixed rates, which simplifies system design but restricts latency and throughput. In this work, we present AERO-VIS, a stereo event-inertial SLAM system with an integrated, data-driven, robust, and performance-optimized keypoint detector. By decoupling event preprocessing from state estimation, our architecture operates asynchronously at the maximum rate supported by available hardware, dynamically adapting to downstream runtime demands and ensuring low-latency, real-time performance. When deploying AERO-VIS on a UAV, we achieve unprecedented accuracy in onboard event-based SLAM. These unique characteristics enable us to present the first purely event-based inertial SLAM system that demonstrates closed-loop UAV control and large-scale state estimation while relying solely on onboard compute. Videos of the experiments, source code, and additional results are available at https://ethz-mrl.github.io/AERO-VIS/.

Yannick Burkhardt, Sebastián Barbas Laina, Simon Boche +2

Mobile Robotics Lab, ETH Zürich · Technical University of Munich · Munich Center for Machine Lerning (MCML) +1

Visual Inertial Odometry Simultaneous Localization And Mapping Vision Sensors Aeroact Real-Time Systems Asynchronous Execution+1

Ask PDF

Jul 2, 2026cs.CV

DL-VINS-Factory: A Modular Framework for Learned Visual Front-Ends in Visual-Inertial SLAM

Deep-learning features excel in visual matching, yet their practical value in tightly coupled visual-inertial SLAM (VI-SLAM) remains insufficiently characterized. We present DL-VINS-Factory, a unified framework that integrates learned feature extractors (ALIKED, RaCo, SuperPoint, XFeat) with either Lucas--Kanade (LK) optical-flow tracking or LightGlue (LG) descriptor matching. All front-ends share a sliding-window Ceres back-end, with optional AnyLoc DINOv2-VLAD loop closure, and 4-DoF pose-graph optimization. We benchmark the system across the four datasets covering indoor, unstructured outdoor, aggressive-motion, and visually degraded conditions. Results show that learned front-ends are viable for real-time embedded VI-SLAM, but are not universally superior to classical tracking. Relative to the corresponding GFTT+LK baseline, ALIKED+LG reduces EuRoC ATE by

5\%

in monocular odometry and by

7\%

in stereo with loop-closure. On NTU-VIRAL, where aggressive aerial motion increases inter-frame viewpoint change, ALIKED+LG stereo reduces loop-closed ATE by

12\%

. In Botanic Garden dataset, optical-flow tracking remains preferable, but learned keypoints still improve over the baseline GFTT, in which SuperPoint+LK reduces grayscale camera ATE by

29\%

, while RaCo+LK reduces RGB camera ATE by

38\%

. On SubT-MRS, learned front-ends display varying degree of improvement based on individual cases. With TensorRT acceleration on a Jetson AGX Orin, all valid configurations run in real time between

29

47

FPS in monocular mode and

18

33

FPS in stereo mode for the EuRoC and NTU-VIRAL datasets. AnyLoc further confirms roughly

2

7\times

more valid loops than BRIEF+DBoW2. The implementation is open-sourced at https://github.com/limshoonkit/DL-VINS-Factory-ROS2/.

Shoon Kit Lim, Melissa Jia Ying Chong, Ting Yang Ling

University of Southampton Malaysia, Iskandar Puteri, Johor, Malaysia · University of Southampton, Southampton SO17 1BJ, U.K.

Orb-Slam2 Simultaneous Localization And Mapping Vision Sensors Optical Flow Modular Framework Web

Ask PDF