The Gap Is in the Geometry: A Convergent Meta-Analysis of Multi-Facet Aerial and Remote Sensing Video Understanding

Georgios Angelidis; Aggelos Psiris; Maria Vakalopoulou; Georgios Th. Papadopoulos; Giorgos Sfikas; Iraklis Varlamis

cs.CVOct 13, 2026

The Gap Is in the Geometry: A Convergent Meta-Analysis of Multi-Facet Aerial and Remote Sensing Video Understanding

Authors: Georgios Angelidis, Aggelos Psiris, Maria Vakalopoulou, Georgios Th. Papadopoulos, Giorgos Sfikas, Iraklis Varlamis

Organizations: Harokopio University of Athens Athens, Greece · CentraleSupelec Paris, France · University of West Attica Athens, Greece

Abstract

Aerial and remote sensing video is increasingly evaluated like web video, by posing multiple-choice questions to multimodal large language models (MLLMs) over sampled frames and reporting accuracy. This paper argues that this recipe measures the wrong aspect, and supports the claim with a convergent meta-analysis of a broad set of published results, organized in four evidence tiers, in which a claim becomes a position only when collected evidence agrees. Parameter scale exhibits a median of +3.4 accuracy points per tenfold increase on aerial benchmarks, far below a human gap of at least 19 points, whereas acquisition conditions control perception, as 315 of 318 evaluations in 19 within-paper contrasts lose accuracy when conditions such as altitude, view, object size or time horizon become harder. No aerial benchmark in the corpus combines video input with per-task human and random baselines, competence collapses from answering to acting in closed-loop navigation, and the same MLLMs score higher on the aerial benchmark than on ground spatial ones, but lower on studies that test cross-view geometry. From these findings the current study derives five falsifiable principles and proposes the AVU-Card, a reporting card for aerial video benchmarks

Figures & tables

Tier	Papers (venue)	Extracted	Normalization
T1 Aerial language-grounded and closed-loop benchmarks (peer-reviewed)	AirCopBench (AAAI’26) [ 59 ] , UrbanVideo-Bench (ACL’25) [ 64 ] , AerialMind (AAAI’26) [ 5 ] , RefDrone (CVPR’26 F) [ 40 ] , Sky2Ground (CVPR’26) [ 45 ] , AeroVerse (TPAMI’26) [ 56 ] , CityEQA (EMNLP’25) [ 65 ] , AerialVLN (ICCV’23) [ 28 ] , CityNav (ICCV’25) [ 23 ] , TravelUAV (ICLR’25) [ 44 ] , UAV-ON (MM’25) [ 51 ] , OpenFly (ICLR’26) [ 18 ]	Per-task tables (24 zero-shot rows on AirCopBench), scale pairs, fine-tuning, closed-loop success, audit	HNS; chance-normalized if no human
T2 Multi-method aerial perception, 2016–2025 (peer-reviewed)	UAV123 [ 32 ] , DTB70 [ 24 ] , UAVDT [ 12 ] , NDFT [ 50 ] , VisDrone DET/VID/SOT/MOT reports and TPAMI [ 13 , 4 , 68 , 14 , 47 , 6 , 69 ] , SAHI [ 1 ] , Okutama [ 2 ] , University-1652 [ 66 ] , ERA [ 31 ] , DroneCrowd [ 46 ] , UAV-Human [ 25 ] , VTUAV [ 63 ] , DroneVehicle [ 39 ] , VISO [ 58 ] , DOTA [ 11 ] , SODA [ 7 ] , SUES-200 [ 70 ] , SeaDronesSee [ 41 ] , MAVREC [ 15 ] , AG-VPReID [ 33 ] , Video2BEV [ 22 ] , CST Anti-UAV [ 52 ]	Per-method scores under two acquisition conditions (5–41 methods per contrast)	Retention $=$ harder / easier, per method; median
T3 Ground and temporal-RS analogues (peer-reviewed)	Spatial benchmarks VSI-Bench [ 54 ] , VLM4D [ 67 ] , SpatialScore [ 49 ] , MMSI-Bench [ 55 ] , Ego3D-Bench [ 20 ] , SPAR-Bench [ 62 ] ; temporal RS benchmarks GEOBench-VLM [ 10 ] , XLRS-Bench [ 42 ] , EarthDial [ 36 ] , TEOChat [ 21 ] , DisasterM3 [ 43 ] , DynamicVL [ 53 ] ; video anchors ATP [ 3 ] , EgoSchema [ 29 ] , Video-MME [ 17 ] , OVO-Bench [ 34 ] , StreamingBench [ 26 ] , TVBench [ 8 ] ; forensic search ForeSea [ 35 ]	Per-task human, chance and blind rows (166 zero-shot rows on the six spatial benchmarks); single- vs multi-temporal and short vs long contrasts	HNS; retention
T4 Preprints 2025–26 ( † )	MM-UAVBench † [ 9 ] , SpatialUAV † [ 61 ] , AeroGround † [ 57 ] , UAV-DualCog † [ 27 ] , UAVReason † [ 37 ] , UAVBench † [ 60 ] , DroneEyes † [ 38 ] , UAVDE-2M † [ 48 ]	Same as T1 (50 zero-shot rows on the three HNS benchmarks)	HNS; used for corroboration only

Table 1: Evidence tiers. 67 papers. Numbers are read from the papers’ own tables; values read from bar charts (UAVDT tracking) or plot legends (VisDrone-SOT19, VTUAV, UAV123) are used only in retention contrasts, never in HNS pooling. Tiers are pooled separately and compared by direction, never averaged together. † marks preprints.

Figure 1: Acquisition conditions control aerial perception (T2). Retention is the score under the harder condition divided by the score under the easier one. Each row is a within-paper contrast; gray dots are individual methods and the blue dot is the median retention. Across the 19 degrading contrasts, 315 of 318 method or configuration evaluations lose accuracy. Orange rows are the two counter-examples in the corpus. UAVDT MOT values are read from bar charts and the VisDrone-SOT19, VTUAV and UAV123 values from plot legends.

Figure 2: Within-family scale slopes on aerial benchmarks (overall accuracy;

\dagger

preprint; preprint family–benchmark contrasts from the three HNS benchmarks). Both tiers give a few points per tenfold increase in parameters, against human gaps of 19–52 points.

Geometry-aware input	Effect	Source
Altitude, view, weather metadata (training)	+12.0 AP night, +11.5 front view, +1.9 high altitude	NDFT [ 50 ]
View-angle experts	+3.1 AP50 (average)	SeaDronesSee [ 41 ]
Camera-motion compensation (homography)	8/8 trackers improve overall	DTB70 [ 24 ]
Geographic semantic map (GNSS + map)	Success $\times$ 2.5–5.7 (test-unseen)	CityNav [ 23 ]
Bird’s-eye-view (3DGS) representation	+4.4 retrieval AP	Video2BEV [ 22 ]
Same-region ground-view pre-training	AP 33.2 $\rightarrow$ 44.8	MAVREC [ 15 ]

Table 2: Geometry-aware inputs and training tested in the corpus (exceptions in the text). Among the aerial question-answering benchmarks, only AeroVerse gives pose to the evaluated models ( Tab. 4 ).

Figure 3: (a) Median-model HNS per competence facet and tier (T1 with 24 models and 13 tasks, T3 with 166 model rows over 6 benchmarks and 56 tasks, T4 with 50 model rows over 3 benchmarks and 43 tasks). T3 perception rests on two tasks. (b) Each point is one model, plotting its HNS on an aerial benchmark against its mean HNS on the ground spatial benchmarks where it is also evaluated. Models score higher on AirCopBench (9/9) and MM-UAVBench † (13/14) than on ground spatial benchmarks, and lower on the geometry-heavy SpatialUAV † and AeroGround † (7 of 8 model–benchmark pairs).

Tier	Temporal contrast	Result	Source
T2	Short $\rightarrow$ long-term SOT, 41 trackers	0.34 (41/41)	[ 14 ]
T2	Short $\rightarrow$ long-term RGB SOT, 18 trackers	0.61 (18/18)	[ 63 ]
T2	30 $\rightarrow$ 10 fps, 14 trackers	0.89 (13/14)	[ 32 ]
T2	Temporal vs. per-frame entries, mean (organizers’ baselines)	15.2 vs. 19.5 (17.7 vs. 15.9) AP	[ 68 ]
T2	Median video vs. median frame model (events)	0.87	[ 31 ]
T3	Single image (mean of 3 sets) $\rightarrow$ multi-temporal fMoW, GPT-4o	84.9 $\rightarrow$ 21.4	[ 36 ]

Table 3: Temporal evidence by tier. Unless stated otherwise, values are retention (harder / easier condition, median over methods; methods that drop in brackets). Margin denotes accuracy above chance.

Reported item	T1 (12)	T4 † (8)
Human baseline, any / per task	5 / 1	4 / 3
Random or chance baseline	7	3
Variance across runs or seeds	0	0
Blind (text-only) score	2	0
Frame-count, clip-length or view-count ablation	4	1
Video as input for all tasks	2	1

Table 4: Reporting audit of the aerial benchmarks in T1 and T4 (counts of benchmarks). The two blind scores come from closed-loop papers (CityEQA, AerialVLN); no aerial question-answering benchmark reports one.

	Claim	T1–3	T4	Status	$\rightarrow$ P
K1	Perception leads the facet profile	✓	✓	Convergent (LOO-sensitive)	P2
K2	Parameter scale is not the lever	✓	✓	Convergent	P1
K3	Acquisition geometry controls accuracy; geometry inputs help	✓	✓	Convergent	P1, P2
K4	Aerial MLLM benchmarks under-test geometry	✓	✓	Divergent, informative	P1, P5
K5	Longer horizons and temporal change cause large losses	✓	✓	Convergent	P3
K6	Answering does not transfer to acting	✓	–	Peer-reviewed only	P3, P5

Table 5: Claim ledger. ✓ supported;

\circ

supported only by ground-view analogues (T3); – not tested in that tier. LOO denotes leave-one-benchmark-out. The last column maps claims to the principles of Sec. 5 .

AVU-Card field	AirCopBench [ 59 ]
1 Facet-coded HNS	14 tasks, 6 facets (13 tasks with human $-$ chance $\geq 20$ ); median-model perception 0.55, multi-view 0.50, ego-motion 0.46, spatial 0.45, planning 0.43, counting 0.11
2 Chance and human	Per task (✓); human raters and subset size not stated in the main text; 1 task with human $-$ chance $=10.3$
3 Blind, single-frame, shuffled floors	Blind models used only to filter questions; no floor reported
4 Variance	Not reported
5 Acquisition envelope	2–6 drones; simulated and real views; altitude, GSD and view angle not reported per item
6 Inputs	Simultaneous still images; no telemetry; frame budget n/a

Table 6: AVU-Card fields with a worked example. Fields 1–4 make accuracy interpretable, 5–7 describe what makes the data aerial, 8–10 bound what the numbers can be used for.

Remote Sensing Benchmarks

Remote sensing benchmarks measure how well models interpret Earth-observation data across sensors, regions, time, tasks, and operating conditions.

See all 193 papers

AirGroundBench: Probing Spatial Intelligence in Multimodal Large Models under Heterogeneous Multi-View Embodied Collaboration
Jun 26, 2026Haotian Li, Yida Wang, Leyuan Wang +7Multi-View Consistency Multimodal Spatial Reasoning
Beijing University of Posts and Telecommunications · University of Electronic Science and Technology of China · Tsinghua University +1
AeroGround: A Comprehensive Benchmark for Aerial-Ground Collaborative Reasoning
Aug 12, 2026Shenghong Yi, Lin Zhang, Muzian Li +6VLM Evaluation VLM Reasoning
Shanghai Innovation Institute · College of Future Information Technology, Fudan University · The Chinese University of Hong Kong +2
ReVA: A Scene-Centric Dataset Beyond Repetition for Remote Sensing Video Question Answering
Sep 28, 2026Zhen Yao, Likai Wang, Yuming Yang +9Remote Sensing VQA Video QA
Lehigh University · University of Southern California · Qualcomm AI Research

Jun 26, 2026J/K move · Enter open · S save

AirGroundBench: Probing Spatial Intelligence in Multimodal Large Models under Heterogeneous Multi-View Embodied Collaboration

Haotian Li, Yida Wang, Leyuan Wang, Jinshan Lai+6

Beijing University of Posts and Telecommunications · University of Electronic Science and Technology of China · Tsinghua University+1

Multi-View Consistency Multimodal Spatial Reasoning Multimodal Reasoning Vision-Language Navigation Multi-Robot Coordination Multimodal Model Evaluation+1

Open paper AskPDF arXiv

In recent years, multimodal large language models (MLLMs) have shown strong potential for embodied intelligence, yet their ability to maintain geometrically consistent spatial understanding across heterogeneous views remains under-evaluated. Existing benchmarks largely focus on single-agent, single-view perception, leaving a gap in the systematic assessment of collaborative air-ground settings, where multi-scale observations are complementary but introduce scale mismatch, asymmetric occlusion, and reference-frame inconsistencies. We present AirGroundBench, a diagnostic benchmark for evaluating multi-view spatial intelligence in heterogeneous UAV-UGV collaboration. AirGroundBench is built from 11 high-fidelity simulated environments with 1,021 synchronized air-ground observation pairs, yielding approximately 62,000 dual-view, four-option single-choice visual question answering instances and 115 closed-loop vision-language navigation episodes. It covers 10 task types organized into four progressively demanding capability dimensions: spatial perception, cross-view alignment, spatial transformation and reasoning, and embodied decision-making. To support geometry-grounded evaluation and analysis, we provide structured spatial annotations, including cross-view object identities and metric 2D and 3D bounding boxes. Evaluations of 13 representative MLLMs under UAV-only, UGV-only, and dual-view input settings reveal consistent bottlenecks: models perform relatively well on spatial perception but struggle with cross-view alignment and transformation-intensive reasoning, and these deficits propagate to sequential decision-making in vision-language navigation. Although dual-view inputs provide measurable gains over single-view variants, a persistent gap from human performance remains, highlighting geometric consistency as a key limitation of current embodied MLLMs.

Abstract

Remote Sensing Benchmarks

Explore similar work

AirGroundBench: Probing Spatial Intelligence in Multimodal Large Models under Heterogeneous Multi-View Embodied Collaboration

AeroGround: A Comprehensive Benchmark for Aerial-Ground Collaborative Reasoning

ReVA: A Scene-Centric Dataset Beyond Repetition for Remote Sensing Video Question Answering