VReID-XFD Challenge: Video ReID Benchmark

• VReID-XFD Challenge is a community benchmark for video-based person re-identification under extreme imaging conditions, testing algorithms across varied aerial and ground views.
• The challenge uses the DetReIDX dataset with 371 identities and rich physical metadata to rigorously evaluate performance via Rank-1, Rank-5, and mAP metrics.
• Results demonstrate that conventional appearance-based methods collapse under severe resolution loss and view variations, driving research toward robust, multi-modal innovations.

The VReID-XFD Challenge is a community benchmark and evaluation campaign for video-based person re-identification (ReID) under extreme far-distance, aerial-to-ground, and ground-to-aerial scenarios. Designed to rigorously test ReID algorithms under severe resolution loss, wide viewpoint changes, unstable motion dynamics, and appearance variability such as clothing changes, VReID-XFD introduces a new operational regime distinct from those of conventional person ReID datasets and protocols. The challenge is centered on the DetReIDX dataset, encompassing multi-platform video sequences, and features rich physical metadata critical for performance analysis and method development. Results reveal systemic accuracy collapses under the strictest imaging conditions, exposing inherent bottlenecks of appearance-based and temporal ReID paradigms, and motivating new research on robustness, invariance, and alternative feature cues (Hambarde et al., 4 Jan 2026).

1. Benchmark Composition and Protocols

VReID-XFD is derived from DetReIDX and comprises 371 unique identities, 11,288 video tracklets, and approximately 11.75 million frames, with acquisition spanning UAV altitudes from 5.8 m to 120 m, pitch angles of 30°, 60°, 90° (nadir), and horizontal distances up to 120 m from targets (Hambarde et al., 4 Jan 2026). The capture procedure consists of two phases: an indoor, high-resolution ground reference session and two outdoor UAV filming sessions with different outfits at seven university campuses. Each tracklet is annotated with physical parameters, including altitude, viewing angle, and distance, as well as 16 soft-biometric attributes.

The evaluation protocol enforces identity-disjoint splits between training and test sets, and supports three scenarios: aerial-to-aerial (A2A, UAV query to UAV gallery), aerial-to-ground (A2G, UAV query to ground gallery), and ground-to-aerial (G2A, ground query to UAV gallery). Challenge participants submit ranked lists of candidate tracklets for each query. Primary performance metrics are the cumulative matching characteristic (CMC) at Rank-1, Rank-5, and Rank-10, and mean Average Precision (mAP), formalized as:

• For query $q$, $N_{gt}$ is the number of ground-truth matches in a gallery of $N$ samples, $P(k)$ is the precision at rank $k$, and $\mathrm{rel}(k) \in \{0,1\}$ signals correct matches:

$$AP_q = \frac{1}{N_{gt}} \sum_{k=1}^{N} P(k)\,\mathrm{rel}(k)$$

$mAP = \frac{1}{Q} \sum_{q=1}^{Q} AP_q$

$$R_k = \frac{1}{Q} \sum_{q=1}^{Q} \mathbf{1}\{\text{true match in top }k\}$$

The dataset and protocols, including detailed annotations and evaluation scripts, are publicly accessible (Hambarde et al., 4 Jan 2026).

2. Challenge Results and Systematic Performance Analysis

The inaugural VReID-XFD-25 Challenge featured 10 international teams and hundreds of submissions. Six teams surpassed the strong VSLA-CLIP baseline, with the top method—SAS-VPReID (DUT_IIAU_LAB)—achieving 43.93% mAP and 37.77% Rank-1 in the most difficult aerial-to-ground scenario. Runners-up included H Nguyn_UIT (39.59% mAP, 33.15% R1) and CJKang’s EAGLE-ReID (39.63% mAP, 33.65% R1) (Hambarde et al., 4 Jan 2026).

Extensive factor analysis, enabled by per-tracklet physical metadata, revealed:

• Altitude: mAP degrades monotonically with increasing UAV height. In A2G, mAP drops from ≈33% (<40 m) to ≈17.7% (>80 m). In A2A, the decline is even sharper: ≈23% to ≈6.6%.
• Horizontal distance: Near-range (10–40 m) yields the highest accuracy; mAP at >80 m is ≈19%. Nadir views with zero offset regain some performance due to maximal resolution.
• Viewing angle: Oblique views (30°) systematically outperform nadir (90°), with a typical mAP gap of ~3–6%.

A universal trade-off is observed: methods with peak accuracy under benign conditions are less stable under extreme degradation, whereas less accurate methods (e.g., JNNCE ISE, H Nguyn_UIT) demonstrate better stability at the highest altitudes. In the hardest regime (high altitude, far distance, nadir), mAP approaches 10–15%, which is marginally above random retrieval.

Key performance degraders include loss of discriminative appearance cues, motion artifacts from UAV instability, and appearance confounds across different sessions (Hambarde et al., 4 Jan 2026).

3. Algorithmic Innovations

A diversity of algorithmic strategies characterize the entries to the VReID-XFD Challenge, with leading approaches addressing the severe degradations through architectural, super-resolution, and adaptive learning mechanisms.

• SAS-VPReID (Yang et al., 9 Jan 2026) integrates:
  • Memory-Enhanced Visual Backbone (MEVB): Combines CLIP ViT-L encoder with multi-proxy memory contrastive learning for robust feature representations.
  • Multi-Granularity Temporal Modeling (MGTM): Processes sequences at multiple temporal scales, using bi-directional “Mamba” operators and learnable fusion, thus capturing both short-term and long-term motion cues.
  • Prior-Regularized Shape Dynamics (PRSD): Extracts body shape dynamics via sequential regression of SMPL shape parameters, regularized to a global prior, making features less sensitive to clothing and resolution loss.
  • These synergistically yield the best known results: A2G mAP = 43.93% (+2.3 over VSLA), G2A mAP = 35.44% (+9.18), A2A mAP = 20.13% (+6.3).
• VSLA-CLIP (Zhang et al., 2024, Endrei et al., 13 Jan 2026): Parameter-efficient adaptation of CLIP vision transformers to video input, featuring:
  • Video Set-Level Adapter (VSLA): Insertion of intra-frame (IFA) and cross-frame (CFAA) adapters for temporal aggregation.
  • Visual–Semantic Alignment: Projecting video features toward CLIP's joint vision–language space via text-based supervision.
  • Platform-Bridge Prompts: Conditioning input to account for platform (ground vs. UAV), explicitly encouraging platform-invariant features.
• S3-CLIP (Endrei et al., 13 Jan 2026): Systematic application of video super-resolution (SwinIR-S backbone) as a preprocessor, trained with task-driven perceptual and temporal consistency losses. This approach significantly improves retrieval in the G→A protocol (R1+11.24%), though gains in A→G are more modest.

These algorithms demonstrate a trend toward robustifying both input and representation; leveraging vision-language modeling, adaptive temporal strategies, and explicit shape/structure priors have proven most successful under VReID-XFD’s strict regime.

4. Limitations, Bottlenecks, and Key Insights

VReID-XFD exposes fundamental limitations of current appearance-based and temporal aggregation ReID systems when confronted with persistent scale, view, and motion degradations (Hambarde et al., 4 Jan 2026). The primary findings include:

• Exclusive reliance on appearance features leads to drastic failure when subjects are rendered as a few pixels, as color and texture information vanish instantaneously with increased altitude.
• Pure temporal models (e.g., extracting gait or motion features) are degraded by platform instability and framewise unpredictability, especially in UAV-acquired sequences.
• Clothing and environmental variations across acquisition sessions, compounded by far-distance uncertainties, further reduce the discriminative utility of existing features.

Systematic analysis confirms that even the highest-performing approach (SAS-VPReID) achieves less than 45% mAP in A2G, and that accuracy collapses are universal in the most challenging conditions. Physical metadata (altitude, angle, distance) are indispensable for both evaluation and algorithmic adaptation.

5. Recommendations and Future Research Directions

Based on challenge outcomes and in-depth analysis, several research directions emerge as promising avenues for advancing video-based person ReID under extreme far-distance and cross-platform conditions (Hambarde et al., 4 Jan 2026):

• Super-resolution and Adaptive Upsampling: Recovering high-frequency details prior to feature extraction is critical in the face of severe scale loss. Task-driven SR networks and adaptive upsampling tailored to anticipated degradations (e.g., via diffusion priors or multi-scale fusion) are prioritized.
• Shape and Gait Priors: Embedding body structure models (e.g., low-rank SMPL parameters, sequence-level shape dynamics) helps retain discriminative capacity when appearance cues are destroyed.
• Multi-modal Fusion: Thermal, depth, or LiDAR sensors provide auxiliary cues potentially robust to view and scale degradation; cross-modal approaches such as the vision-RF pipeline (Cao et al., 2022) have shown high accuracy and robustness in heterogeneous environments.
• Meta-learning and Adaptive Conditioning: Incorporating altitude, angle, and distance as explicit conditioning variables or during domain adaptation could enhance model invariance.
• Self-supervised Temporal Aggregation: Robust temporal modeling methods that can tolerate platform-induced noise and instability.

The VReID-XFD dataset and challenge infrastructure aim to provide a testbed for these future approaches, enabling systematic measurement of progress against the unique challenges presented by extreme far-distance, cross-platform ReID.

6. Dataset Impact and Community Adoption

VReID-XFD represents a significant advance in the evaluation of person ReID under realities of urban-scale surveillance and aerial-ground coordination. Its unique combination of rich physical metadata, tracklet-level evaluation, and harsh imaging regimes has established new lower bounds for achievable accuracy and set the research agenda for robustness, invariance, and multimodality (Hambarde et al., 4 Jan 2026). The dataset’s public availability, detailed annotation, and open protocols expedite reproducibility and cross-method comparison. Early results from the challenge indicate it is already catalyzing methodological innovation, with direct extensions and benchmarking in subsequent large-scale projects such as G2A-VReID (Zhang et al., 2024) and super-resolution-aided frameworks (Endrei et al., 13 Jan 2026).

7. Comparative and Related Efforts

Complementary lines of research build on insights from VReID-XFD:

• VSLA-CLIP and G2A-VReID (Zhang et al., 2024) established a baseline for visual-semantic alignment and platform-adaptive ReID using CLIP, with results informing and subsequently surpassed by VReID-XFD entrants.
• S3-CLIP (Endrei et al., 13 Jan 2026) is the first systematic investigation into video super-resolution’s utility for ReID, specifically under challenging VReID-XFD settings, reporting notable rank improvement in ground-to-aerial retrieval.
• Sensor Fusion Methods: The vision–RF gait ReID approach (Cao et al., 2022) demonstrates robustness, privacy consideration, and cross-modal identification, achieving ≈93% top-1 accuracy in more moderate but cross-modal settings.
• Vehicle ReID in Fisheye and Surround-view (Wu et al., 2020): Although problem domain differs (vehicle, not person), the solution architecture—incorporating spatial constraints, attention-based feature extraction, and drift-aware tracking—maps directly onto future extensions of VReID-XFD methodology.

VReID-XFD has thus become a focal point for benchmarking and stimulating innovation in next-generation video-based ReID systems for adverse and real-world operating environments.