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

1. Introduction

• The central position is that the aerial-video understanding gap lies in geometry and time, while current benchmarks are not built to measure it.A benchmark can report steady progress while leaving operator-relevant capabilities untested.Source: §1 · p. 1
• The analysis investigates what 67 published works imply by comparing within-paper contrasts and normalizing results against each paper’s chance and human baselines.Source: §1 · p. 1
• The paper contributes a tiered meta-analysis, eleven graded claims, five falsifiable principles, and the AVU-Card reporting card for aerial-video benchmarks.The meta-analysis includes human-normalized scores for 112 tasks and 240 zero-shot model rows in 10 MLLM benchmarks, 21 retention contrasts across 333 evaluations, model-matched aerial–ground comparisons, and an audit of 20 benchmarks.Source: §1 · p. 2

2. What Makes Aerial Video Different

• Aerial video understanding is multi-facet across seven competence areas and acquisition conditions including altitude, view angle, object size, time horizon, frame rate, illumination, sensor, and platform count.The seven areas are perception, counting, spatial and temporal reasoning, multi-view association, ego-motion or self-state, and planning.Source: §2 · p. 2
• Aerial footage spans drone video at 10–60 fps, satellite video at a few to 30 fps, and revisits separated by days, months, or years.Benchmarks in the corpus typically cover one temporal scale, from a single instant to flight clips lasting 10 seconds to 10 minutes.Source: §2 · p. 2
• Unlike web-video benchmarks, aerial benchmarks vary acquisition conditions that affect perception, including camera motion, altitude-linked ground sampling distance, and geometry across platforms.Objects of interest may be only a few pixels wide, and relevant views must be related through geometry rather than appearance.Source: §2 · p. 2
• Prior work spans aerial perception, language-grounded benchmarks, aerial MLLM video tasks, temporal remote-sensing benchmarks, and ground-view spatial-intelligence studies.The paper pools these literatures quantitatively to ask what aerial video benchmarks must report for accuracy claims to be supported.Source: §2 · p. 2

3. Meta-Analysis Method

• The corpus includes papers reporting numbers relevant to at least one analysis and treats papers without a verifiable peer-reviewed version as preprints.Table 1 covers 67 papers; tiers are pooled separately and compared by direction, never averaged together.Source: §3 · p. 2, Table 1 · p. 3
• Human-normalized scores scale model performance relative to each task’s own chance and human accuracy, with 0 representing chance and 1 human level.Negative values are retained; tasks with a human–chance gap below 20 points are excluded as unstable.Source: §3 · p. 2
• Facet scores average a model’s task-level human-normalized scores, while benchmark and tier facet values use medians across models or benchmark rows.Tasks are coded to one of seven competence facets.Source: §3 · p. 3
• Retention contrasts compare the same methods under easier and harder acquisition conditions by dividing the harder-condition score by the easier-condition score.The analysis reports median retention and the number of methods losing accuracy, while retaining contrary contrasts as counter-examples.Source: §3 · p. 3
• Model-matched comparisons use each model’s mean human-normalized score across all tasks of an aerial benchmark and a ground spatial benchmark.Source: §3 · p. 3
• A claim is convergent when peer-reviewed and preprint evidence agree in direction; sensitivity checks compare best-model with median-model scores and leave one benchmark out per tier.Claims supported only by preprints are tentative, while those supported only by pre-MLLM studies are established for classical perception with transfer to MLLMs left open.Source: §3 · p. 3

4.1. Scale, appearance and geometry

• +3.4 overall accuracy points per tenfold increase in parameters is the median scale gain on peer-reviewed aerial benchmarks, versus a 19.1-point best-open-model-to-human gap on AirCopBench.The preprint HNS benchmarks show a median +5.8 points across eight family–benchmark contrasts.Source: §4.1 · p. 3, Figure 2 · p. 4
• Perception leads the typical model’s facet profile, but the ranking is sensitive to the best-model comparison and MM-UAVBench†’s contribution to pooled preprints.AirCopBench median-model perception is 0.55 versus 0.11–0.50 on other facets; the corresponding T4 figures are 0.60 versus 0.13–0.33.Source: §4.1 · p. 3, Figure 3 · p. 5
• 315 of 318 evaluations lose accuracy across 19 degrading acquisition contrasts, with altitude, view, object size, and geometry shaping the losses.The counter-examples show altitude is task-dependent: on SeaDronesSee the highest band is easiest, while night conditions differ for detection and tracking.Source: §4.1 · p. 4, Figure 1 · p. 4
• Models score higher on AirCopBench and MM-UAVBench† than on ground spatial benchmarks, but lower on geometry-heavy SpatialUAV† and AeroGround† in 7 of 8 pairs.The median differences are +0.27 and +0.19 HNS on AirCopBench and MM-UAVBench†, versus −0.08 and −0.15 on SpatialUAV† and AeroGround†.Source: §4.1 · pp. 4–5, Figure 3 · p. 5

4.2. Time, action and evaluation

• Long horizons are the largest loss for classical aerial trackers, while reduced frame rate harms 13 of 14 and MLLM temporal scores lack frame-count or order ablations.On TEOChat’s QFabric task, all five models gain from two to five images; reverse-sequence performance on AeroGround† has median HNS 0.04.Source: §4.2 · p. 5, Table 3 · p. 6
• Closed-loop aerial navigation scores are low and generalization to unseen settings further reduces success, despite stronger performance when an MLLM receives the right image.On CityEQA, the MLLM scores 4.37/5 with the right image, while the full agent reaches 3.00 against 4.94 for humans.Source: §4.2 · pp. 5–6
• No peer-reviewed aerial benchmark combines video input with per-task human and random baselines, and none reports variance across runs.Only AirCopBench reports per-task human and random baselines among peer-reviewed benchmarks, but it uses still images.Source: §4.2 · p. 6, Table 4 · p. 6
• Geometry-aware view comparisons and ego-motion evidence vary by setting, while the human gap and persistent pedestrian difficulty remain visible across the corpus.Across thresholds, 8 or 9 of 9 AirCopBench models score higher than on ground benchmarks, and 7 of 8 cross-view preprint pairs score lower.Source: §4.2 · pp. 6–7, Table 5 · p. 7

P1. Geometry as an input

• P1 proposes camera pose, altitude, and ground sampling distance as model inputs, testable by whether telemetry improves ego-motion and multi-view tasks at equal training data.The corpus mostly finds gains from geometry-aware inputs, while parameter scale moves accuracy little and fine-tuning barely changes the peer-reviewed self-state task.Source: §5 · p. 7

P2. Query-driven resolution

• P2 proposes allocating token or pixel budgets by query and ground sampling distance, with under-10-pixel object recall at equal token count as the falsification test.Perception leads median aerial facet scores, while small and distant targets lose one to three quarters of accuracy.Source: §5 · p. 7

P3. Time as a continuum

• P3 calls for streaming, forward-looking questions across at least two time scales, with models retaining state and full-model performance tested against single- or shuffled-frame baselines.Long horizons are among the corpus’s largest losses, including 0.34 retention across 41 trackers.Source: §5 · p. 7

P4. Platforms as a team

• P4 proposes shared geo-registered memory for multiple-drone and air–ground views, falsified if naive image concatenation matches geo-registered fusion on cross-view tasks.The principle responds to appearance-based association failing across large viewpoint changes.Source: §5 · p. 7

P5. Evaluate the construct

• P5 requires benchmarks to show their questions test the named capability through human and chance baselines, blind and single-frame floors, variance, evidence localization, and closed-loop transfer when action is implied.The principle is falsified if benchmarks with these controls rank models no differently from those without them.Source: §5 · p. 7

What the fields capture

• Fields 1–4 help determine whether an accuracy gain reflects the named capability, while fields 5–7 record acquisition conditions, telemetry, and time scales.The card is designed to use quantities benchmarks already compute or record, without new annotation beyond a small human study.Source: §6 · p. 7

Card structure

• The AVU-Card’s ten fields cover accuracy interpretation, aerial data conditions, and limits on how results can be used.The table presents the fields with a worked example.Source: §6 · p. 7, Table 6 · p. 8

Reporting gaps and use boundaries

• Fields 8–10 record evidence ties, closed-loop transfer, and whether tasks identify people; the best-documented peer-reviewed aerial MLLM benchmark leaves fields 3, 4, and 8 empty and fields 2 and 5 incomplete.The authors describe this as evidence about field reporting norms, not a judgment about that benchmark.Source: §6 · pp. 7–8

Alternative views

• Observed scale slopes imply that parameter scale alone would need several orders of magnitude to close the gap, but the evidence does not rule out larger or better-trained models doing so.Fine-tuning gains are uneven across facets, and the peer-reviewed self-state task barely moves; explicit 3D structure narrows gaps on Ego3D-Bench.Source: §7 · p. 8
• The position allows modular systems, but still calls for testing composed systems on geometric questions; explicit 3D maps substantially improve Ego3D-Bench scores.An estimated cognitive map raises InternVL3-8B from 43.1 to 56.0, and a ground-truth map to 79.4, against 85.3 for humans [20].Source: §7 · p. 8

Limitations

• Because only published numbers are used, blind, single-frame, and shuffled-frame floors could not be measured on aerial benchmarks and remain a key test for future work.The analysis also inherits source-paper choices, including subset or limited-rater human baselines and some small-subset contrasts.Source: §7 · p. 8
• The model-matched comparison supports the direction of the aerial-versus-ground finding, not the size of differences, because prompts, frame sampling, and answer formats vary.For seven of eight cross-view pairs, the ground side is a single benchmark.Source: §7 · p. 8

Implications and governance

• Forensic search should be scored on returned evidence as well as answers, and the paper recommends flagging identity-related tasks amid strict legal limits on aerial video of people.In the corpus, identity is an explicit target only in re-identification and person-search tasks [25] [33] [35]; no aerial MLLM benchmark flags identifiable-person questions.Source: §7 · p. 8

• Appearance is the strongest facet for typical models, while accuracy is controlled by altitude, view, object size, and time horizon.Source: §8 · p. 9
• Scale buys little, and competence does not survive the step from answering questions to acting.Source: §8 · p. 9
• Current benchmarks cannot yet show whether their questions require video or geometry.Source: §8 · p. 9
• The paper calls for models informed about camera position and benchmarks documented with the AVU-Card.Source: §8 · p. 9