A computer-vision system designed to identify the same person in ground-level and aerial images scored higher than the strongest listed baseline, but the task remained difficult. On the test set, the proposed ICRG framework recorded a pair F1 of 22.28%, compared with 16.65% for GroundingDINO-T, the strongest of seven single-view baselines. Pair F1 is the paper's measure of how well the system selects a matching person in both views, rather than a plausible box in only one.
Only 12.54% of test samples had both views meet the paper's intersection-over-union threshold of 0.5. Intersection-over-union, or IoU, compares the overlap between a predicted person box and the correct box. The result shows that paired detection is harder than obtaining plausible boxes from individual views.
A test built around the same identity
The work introduces A-PAIR, a cross-view benchmark derived from the G2APS air-ground person dataset. It contains 22,137 cross-view referring samples over 7,588 unique ground images and 3,891 unique aerial images. The central task is to ground one natural-language referring expression to the same physical person in the paired views.
Those samples are split into 15,497 training samples, 2,213 validation samples and 4,427 test samples. The experiments use that A-PAIR protocol, and the comparison with the seven single-view baselines is made on an image-disjoint test split. The test is therefore centered on agreement between the two views, not just a plausible detection in one image.
To build the supervision, the paper describes FARA, a semi-automatic annotation framework that generates factorized referring descriptions and identity-consistency supervision. Its resulting training export contains 377,402 pairs, including 50,741 positive pairs. That gives the model language information about the target alongside a signal about whether the views refer to the same identity.
What the components added
ICRG combines factorized referential grounding, candidate-completeness supervision and cross-view consistency calibration for joint air-ground pair selection. The reported ICRG scores were 35.17% for ground instance-level F1, 21.89% for aerial instance-level F1 and 22.28% for pair F1. The last figure captures the added requirement that the selected detections match across views.
The ablation results show a large gap between using the factorized signal and using none of the listed signals. With Fact. alone, pair F1 was 20.88%, compared with 3.60% for the no-signal configuration. These are comparisons between model settings, so they should be read as benchmark results rather than proof that Fact. alone caused the final performance.
Adding candidate-completeness supervision to Fact. corresponded to higher aerial scores. Aerial instance-level F1 rose from 20.83% to 21.82%, and aerial image-level F1 from 34.47% to 35.82%; the full model reached 21.89% and 35.92% on those measures. The figures describe observed differences between the tested configurations.
Consistency calibration made a smaller change in the reported pair measures. Adding Cons. to Fact.+Cand. changed pair accuracy from 12.24% to 12.54% and pair F1 from 21.82% to 22.28%. The authors interpret this pattern as support for explicit cross-view consistency modeling, but the ablation does not isolate a causal effect.
A benchmark result, not a deployment test
The challenge is uneven across views. The ICRG row reports ground instance-level F1 of 35.17%, versus 21.89% for aerial instances. Alongside the low paired-hit rate, that gap underlines how much harder it is to make the two detections agree than to find a plausible person in a single view.
These findings come from a preprint benchmark evaluation using the A-PAIR protocol. Because A-PAIR is derived from G2APS, the reported results establish performance on this data source, but they do not establish how well the framework will transfer to other paired air-ground data.
The evidence supports a comparative performance claim within A-PAIR, not causal superiority for ICRG or any one component. Independent evaluation on other paired air-ground data is needed to assess generalization beyond the benchmark.
Paper data and sources
Original title: A-PAIR: A Benchmark and Identity-Consistent Grounding Framework for Air-Ground Cross-View Referring Person Detection
Authors: Zhoupeng Guo, Xinjie Yao, Yunqi Zhu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text