M3GA-Wild: A Large-Scale Dataset and Benchmark for Multi-Modal Multi-session Ground-to-Aerial Place Recognition in Forests

Ethan Griffiths; Maryam Haghighat; Simon Denman; Clinton Fookes; Milad Ramezani

cs.CVSep 19, 2026

M3GA-Wild: A Large-Scale Dataset and Benchmark for Multi-Modal Multi-session Ground-to-Aerial Place Recognition in Forests

Authors: Ethan Griffiths, Maryam Haghighat, Simon Denman, Clinton Fookes, Milad Ramezani

Organizations: Queensland University of Technology (QUT) · CSIRO Robotics

Abstract

We present M3GA-Wild, the first benchmark for multi-modal, multi-session ground-to-aerial place recognition in forests. M3GA-Wild unifies and extends existing forest localisation datasets, providing a holistic benchmark with synchronised RGB imagery and LiDAR from ground traversals spanning 36 km, aligned high-resolution aerial imagery and multi-altitude LiDAR covering 370 hectares, and accurate geo-referenced 6-DoF poses for precise evaluation. M3GA-Wild captures diverse forest scenes with varying viewpoints, occlusion, and environmental conditions, enabling systematic evaluation of visual, LiDAR, cross-modal, and multi-modal methods. Baseline experiments show that LiDAR-based approaches significantly outperform vision-only methods under severe viewpoint differences, while current multi-modal fusion strategies yield limited gains due to poor cross-modal alignment. By pairing aerial RGB imagery with geo-referenced aerial LiDAR, M3GA-Wild also enables evaluation of foundation models for monocular depth estimation as a cheap source of 3D geometry from forest imagery, with initial experiments revealing shortfalls of current methods. These results highlight key challenges in cross-platform localisation, including modality misalignment and severe domain gaps. M3GA-Wild establishes a new benchmark to support research in robust multi-modal localisation and long-term autonomy in unstructured natural environments. The dataset and code will be available upon acceptance.

Figures & tables

Dataset	Year	RGB	LiDAR	Ground	Aerial	Environment	Scale (Ground / Aerial)
KITTI [ 13 ]	2013	✓	✓	✓	✗	Urban	44 km
Oxford RobotCar [ 32 ]	2017	✓	✓	✓	✗	Urban	1000 km
BotanicGarden [ 31 ]	2020	✓	✓	✓	✗	Outdoor	17 km
MulRan [ 22 ]	2020	✗	✓	✓	✗	Urban	123 km
Wild-Places [ 24 ]	2023	✗	✓	✓	✗	Forest	33 km
MARS-LVIG [ 30 ]	2024	✓	✓	✗	✓	Urban / Outdoor	1.2 km 2

Table 1: Comparison of existing place recognition and localisation datasets with LiDAR data. Many existing datasets focus on ground-to-ground place recognition. M3GA-Wild is the first dataset to consider multi-modal ground-to-aerial place recognition in forests.

Forest	Images				LiDAR Submaps
	Train	Val	Test		Train	Val	Test
	Train	Val	Query	DB	Train	Val	Query	DB
Karawatha	231.8 k	6.8 k	11.2 k	17.8 k	37.3 k	2.5 k	7.4 k	17.8 k
Venman	124.4 k	–	9.6 k	12.4 k	26.7 k	–	6.4 k	12.4 k
QCAT	–	–	–	0.4 k	–	–	0.8 k	0.4 k
Samford	–	–	–	4.6 k	–	–	1.3 k	4.6 k

Table 2: Dataset statistics for each forest region of M3GA-Wild.

	Method	Re-Ranker	Karawatha			Venman			QCAT			Samford			Average
	Method	Re-Ranker	R1	R5	R10	R1	R5	R10	R1	R5	R10	R1	R5	R10	R1	R5	R10
Low Altitude (LA)	MinkLoc3Dv2 [ 28 ]	—	46.2	63.0	70.1	63.6	82.7	90.5	49.5	73.0	83.8	34.4	53.1	61.2	48.5	67.9	76.4
	EgoNN [ 26 ]	—	45.4	63.4	73.1	56.4	78.8	86.7	42.1	75.4	85.8	49.3	68.2	76.8	48.3	71.4	80.6
	LoGG3D-Net [ 43 ]	—	21.7	37.2	45.5	27.7	43.1	52.1	31.1	50.3	56.8	7.3	12.2	15.5	21.9	35.7	42.5
	CrossLoc3D [ 18 ]	—	7.9	20.0	26.4	11.7	28.1	38.6	16.5	37.9	55.8	4.0	11.7	16.0	10.0	24.4	34.2
	HOTFormerLoc [ 17 ]	—	50.2	65.5	72.7	50.8	67.8	76.1	62.8	83.1	92.3	73.8	85.1	90.3	59.4	75.4	82.8

Table 3: LiDAR place recognition results on M3GA-Wild, for low and high altitude LiDAR data.

	Method	Karawatha			Venman			Average
	Method	R1	R5	R10	R1	R5	R10	R1	R5	R10
Zero-shot	NetVLAD [ 4 ]	0.40	1.01	1.79	0.27	1.24	2.39	0.33	1.12	2.09
	MixVPR [ 2 ]	0.60	1.92	3.23	0.14	0.49	1.14	0.37	1.21	2.18
	SALAD [ 19 ]	0.54	1.64	2.92	0.25	0.73	1.42	0.40	1.18	2.17
	BoQ [ 3 ]	0.92	2.67	3.85	0.08	0.36	0.79	0.50	1.51	2.32
	AnyLoc [ 21 ]	2.11	4.43	6.10	0.72	2.51	4.52	1.41	3.47	5.31

Table 4: Visual place recognition results on M3GA-Wild.

Method (Image Encoder)	Modality		Karawatha			Venman			QCAT			Samford			Average
Method (Image Encoder)	Ground	Aerial	R1	R5	R10	R1	R5	R10	R1	R5	R10	R1	R5	R10	R1	R5	R10
Lip-Loc [ 38 ] (ResNet50)	Range Img	RGB	1.32	5.27	8.28	5.17	11.05	14.59	4.13	14.32	19.23	3.22	6.37	9.36	3.46	9.25	12.87
Lip-Loc [ 38 ] (DINOv2)	Range Img	RGB	1.72	4.38	7.47	0.97	3.03	4.46	17.29	40.39	53.03	0.75	2.10	3.15	5.18	12.48	17.03
Lip-Loc [ 38 ] (DINOv3)	Range Img	RGB	1.63	3.29	5.14	2.16	5.85	8.67	17.42	34.58	43.35	0.00	0.00	0.52	5.30	10.93	14.42
Lip-Loc [ 38 ] (ResNet50)	Height Map	RGB	2.23	6.85	11.05	3.30	8.13	11.65	5.29	20.77	29.16	0.60	1.65	3.37	2.86	9.35	13.81
Lip-Loc [ 38 ] (DINOv2)	Height Map	RGB	1.13	3.96	6.41	0.79	1.90	3.04	8.13	26.58	38.45	0.75	2.85	4.19	2.70	8.82	13.02

Table 5: Cross-modal place recognition results on M3GA-Wild.

Method	Karawatha			Venman			Average
Method	R1	R5	R10	R1	R5	R10	R1	R5	R10
MinkLoc++ [ 25 ] (RN18)	29.4	48.3	59.3	29.0	51.1	62.0	29.2	49.7	60.7
MinkLoc++v2 (RN50)	48.2	69.7	77.4	37.7	60.7	71.0	43.0	65.2	74.2
MinkLoc++v2 (DINOv2)	46.7	68.6	77.1	48.9	76.8	86.8	47.8	72.7	81.9
UMF [ 11 ] (RN50)	24.7	43.2	52.0	12.9	29.4	41.1	18.8	36.3	46.6
∗ RN means ResNet.

Table 6: Multi-modal place recognition results on M3GA-Wild.

Method	GNSS Limit	Image-Only			LiDAR-Only			Multi-Modal
Method	GNSS Limit	R1	R5	R10	R1	R5	R10	R1	R5	R10
MLoc++ [ 25 ] (RN18)	–	1.0	2.4	3.4	36.2	58.5	69.5	29.2	49.7	60.7
MLoc++v2 (RN50)	–	0.6	1.9	3.6	50.4	72.1	80.0	43.0	65.2	74.2
MLoc++v2 (DINOv2)	–	1.0	4.0	7.0	51.6	75.4	83.1	47.8	72.7	81.9
MLoc++v2 (DINOv2) ‡	–	4.0	8.9	12.5	35.9	59.7	70.4	41.7	65.8	76.0
MLoc++v2 (DINOv2)	100 m	11.1	33.0	48.8	73.6	91.6	95.7	70.9	90.0	95.1

Table 7: Performance analysis of each modality branch in MinkLoc++ (MLoc++), with pseudo-GNSS constraints on the query search space. Average results over all sequences are reported.

Panel of Figure 2 — Figure 2: Queries where MinkLoc++v2 (DINOv2) image-only branch successfully localised where LiDAR-only did not. RGB captures ground texture details that are invisible to LiDAR, demonstrating that RGB can provide useful information in forests if integrated correctly.

	Model	LR	Epoch	BS	Res (W, H)	Other
LPR	MinkLoc3Dv2 [ 28 ]	$1\mathrm{e}^{-3}$	120	2048	–
	EgoNN [ 26 ]	$1\mathrm{e}^{-3}$	120	128	–
	LoGG3D-Net [ 43 ]	$1\mathrm{e}^{-3}$	50	12	–	Feat dim: 32
	CrossLoc3D [ 18 ]	$5\mathrm{e}^{-3}$	200	128	–
	HOTFormerLoc [ 17 ]	$8\mathrm{e}^{-4}$	100	2048	–
	HOTFLoc++ [ 16 ]	$8\mathrm{e}^{-4}$	60	256	–

Table S1 : Model-specific hyperparameters used on M3GA-Wild.

Forest	Sequence	Coverage	# Images	# Submaps
Karawatha	2019_06_11_airborne ∗	1.80 km 2	17792 ∗	17792 ∗
	2021_06_21_K-02_ground	5.66 km	75570	10075
	2021_06_22_K-01_ground	5.14 km	66120	8816
	2021_12_13_K-03_ground	6.27 km	14205	15150
	2022_08_11_K-04_ground	2.81 km	43766	5805
	2024_07_10_aerial ∗	1.80 km 2	17792 ∗	17792

Table S2 : Summary of M3GA-Wild sequences.

Panel of Figure 11 — Figure S1 : Comparison of CHMv2 [ 6 ] height maps vs . LA aerial LiDAR ground truth. From top to bottom: Karawatha, Venman, QCAT, Samford.

Panel of Figure 12 — Figure S2 : Comparison of low and high altitude aerial LiDAR submaps in M3GA-Wild, from the ground camera perspective. The poor beam penetration in very dense canopy exacerbates the ground–aerial matching problem.

Panel of Figure 13 — Figure S3 : Overview of each forest in M3GA-Wild. Aerial RGB imagery is overlaid with ground trajectory (blue), test regions (red), and high altitude LiDAR scans colourised by z coordinate.

Panel of Figure 14 — Figure S4 : Average Recall@N curves for each forest in M3GA-Wild, comparing LPR, VPR, and MMPR models for low altitude LiDAR.

Visual Place Recognition

Visual Place Recognition matches camera observations to mapped places despite changes in viewpoint, lighting, weather, or time.

See all 84 papers

Viking Hill Dataset: A Lidar-Radar-Camera Dataset for Detection and Segmentation in Forest Scenes
Jun 17, 2026Vladimír Kubelka, Oleksandr Kotlyar, Unal Artan +1Millimeter-Wave Radar mmWave Radar Sensing
Örebro University, AASS research centre, Robot Navigation and Perception Lab, Örebro, Sweden
WildCross: A Cross-Modal Large Scale Benchmark for Place Recognition and Metric Depth Estimation in Natural Environments
Mar 2, 2026Joshua Knights, Joseph Reid, Kaushik Roy +3Visual Place Recognition Metric Depth Estimation
CSIRO Robotics, Data61, CSIRO, Australia. · Queensland University of Technology (QUT), Australia.
MAG-VLAQ: Multi-modal Aerial-Ground Query Aggregation for Cross-View Place Recognition
May 10, 2026Zhengyi Xu, Yuhang Ming, Zhihao Zhan +3Cross-View Geo-Localization Visual Place Recognition
Hangzhou Dianzi University · TopXGun Robotics · University of Zaragoza

Jun 17, 2026J/K move · Enter open · S save

Viking Hill Dataset: A Lidar-Radar-Camera Dataset for Detection and Segmentation in Forest Scenes

Vladimír Kubelka, Oleksandr Kotlyar, Unal Artan, Martin Magnusson

Örebro University, AASS research centre, Robot Navigation and Perception Lab, Örebro, Sweden

Millimeter-Wave Radar mmWave Radar Sensing 3D Semantic Segmentation Robotic Perception Multimodal Sensor Fusion

Open paper AskPDF arXiv

Autonomous robots operating under forest canopies need robust perception of trees and surrounding vegetation across varying seasonal conditions. Existing forestry datasets provide lidar or camera data with per-tree annotations, but none include co-registered 4D imaging radar -- a modality of growing interest for its resilience to visual degradation, surface contamination, and vegetation occlusion. We introduce a multi-sensor forest dataset collected by a mobile robot equipped with a high-resolution FMCW imaging radar, lidar, RGB camera, IMU, and RTK-GNSS. The site was recorded in two sessions under contrasting vegetation states, and 3D cuboid annotations -- including per-tree diameter estimates -- provide shared semantic labels across all three perception modalities. Furthermore, we provide baseline results for semantic segmentation of the radar and lidar point clouds using MinkowskiUNet. Radar achieves IoU scores competitive with lidar for dominant classes (ground 91%, canopy 86%) while lagging on geometrically fine structures such as tree trunks (56% vs. 74%). A cross-modality analysis further compares lidar and radar trunk segmentation against an RGB detection model, and a diameter-stratified evaluation reveals how trunk segmentation quality varies with tree size. Beyond segmentation, the co-registered multi-modal data and RTK-GNSS-aided reference positioning support research in mapping, localization, and sensor fusion under canopy. The dataset and annotation tools are publicly available.

Abstract

Visual Place Recognition

Explore similar work

Viking Hill Dataset: A Lidar-Radar-Camera Dataset for Detection and Segmentation in Forest Scenes

WildCross: A Cross-Modal Large Scale Benchmark for Place Recognition and Metric Depth Estimation in Natural Environments

MAG-VLAQ: Multi-modal Aerial-Ground Query Aggregation for Cross-View Place Recognition