VReID-XFD: Video-based Person Re-identification at Extreme Far Distance Challenge Results

TL;DR

This work introduces VReID-XFD, a video-based benchmark and challenge for extreme far-distance aerial–ground person re-identification, derived from DetReIDX and spanning altitudes 5.8–120 m, angles 30°–90°, and distances up to 120 m. The dataset includes 371 identities, 11,288 tracklets, and 11.75 million frames across seven campuses, with three strict cross-view evaluation protocols (A→A, A→G, G→A) and rich metadata. A Kaggle-based evaluation with ten teams reveals a persistent performance decline with altitude, distance, and nadir viewing, with the best method achieving 43.93% mAP in A→G, highlighting the substantial challenge of extreme-distance ReID. The study provides a rigorous testbed for assessing cross-view temporal representations and informs future directions toward robust, operation-ready aerial surveillance systems. The dataset and evaluation protocols are publicly available to enable ongoing research and benchmarking.

Abstract

Person re-identification (ReID) across aerial and ground views at extreme far distances introduces a distinct operating regime where severe resolution degradation, extreme viewpoint changes, unstable motion cues, and clothing variation jointly undermine the appearance-based assumptions of existing ReID systems. To study this regime, we introduce VReID-XFD, a video-based benchmark and community challenge for extreme far-distance (XFD) aerial-to-ground person re-identification. VReID-XFD is derived from the DetReIDX dataset and comprises 371 identities, 11,288 tracklets, and 11.75 million frames, captured across altitudes from 5.8 m to 120 m, viewing angles from oblique (30 degrees) to nadir (90 degrees), and horizontal distances up to 120 m. The benchmark supports aerial-to-aerial, aerial-to-ground, and ground-to-aerial evaluation under strict identity-disjoint splits, with rich physical metadata. The VReID-XFD-25 Challenge attracted 10 teams with hundreds of submissions. Systematic analysis reveals monotonic performance degradation with altitude and distance, a universal disadvantage of nadir views, and a trade-off between peak performance and robustness. Even the best-performing SAS-PReID method achieves only 43.93 percent mAP in the aerial-to-ground setting. The dataset, annotations, and official evaluation protocols are publicly available at https://www.it.ubi.pt/DetReIDX/ .

Figures (9)

• Figure 1: Overview of the VReID-XFD data collection and evaluation protocol. The dataset adopts a two-phase acquisition process with indoor ground reference capture and outdoor UAV-based extreme-distance observation, enabling aerial--ground and aerial--aerial cross-view evaluation.
• Figure 2: Distribution of annotated soft-biometric attributes in VReID-XFD, including appearance, clothing, and physical traits used for analysis and evaluation.
• Figure 3: Scatter plots of ROIs height/width in three different distance bins. The bottom-right plot compares the distribution of the ROIs height (in pixels) in indoor and outdoor data.
• Figure 4: Aerial→Ground test case at high altitude and top angle. Green/red: correct/incorrect labels.
• Figure 5: Comprehensive analysis of physical factors affecting video-based person re-identification performance under extreme far-distance conditions. (a) Effect of altitude, (b) effect of viewing angle, and (c) effect of horizontal distance. All methods exhibit monotonic degradation under increasingly adverse imaging conditions.