DetReIDX Leaderboard

Track state-of-the-art performance on the DetReIDX dataset across multiple computer vision tasks.

Person Detection

Task Description

The person detection task evaluates a model's ability to accurately locate and identify humans in drone-captured imagery across varying altitudes (5-120m), distances (10-120m), and viewing angles (30°, 60°, 90°).

Metrics
  • AP50: Average Precision at IoU threshold of 0.5
  • AP75: Average Precision at IoU threshold of 0.75
  • mAP: Mean Average Precision across IoU thresholds from 0.5 to 0.95 with a step size of 0.05
Experimental Setup

All models are evaluated using the official DetReIDX test split containing 108,252 images with 4,217,824 bounding box annotations across various environmental conditions and viewpoints.

# Method AP50 AP75 mAP D1 (≤20m) D2 (20-50m) D3 (≥50m) Paper/Code Submitted
1 YOLOv8-L CNN 0.734 0.495 0.531 0.885 0.688 0.137 Paper Code 2025-03-29
2 DDOD-R50 CNN 0.608 0.432 0.487 0.857 0.776 0.111 Paper 2025-03-20
3 Grid-RCNN CNN 0.620 0.440 0.492 0.839 0.770 0.150 Paper 2025-03-15

Person Re-Identification

Task Description

The person re-identification task evaluates a model's ability to match individuals across different viewpoints, sessions, and camera types. DetReIDX offers three challenging scenarios:

  • Aerial→Aerial (A2A): Matching between different UAV viewpoints, including across different altitudes and days (with clothing changes)
  • Aerial→Ground (A2G): Matching UAV captures to ground-level imagery
  • Ground→Aerial (G2A): Matching ground-level images to UAV captures
Metrics
  • mAP: Mean Average Precision
  • Rank-1/5/10: Cumulative Matching Characteristic at Ranks 1, 5, and 10
Dataset Statistics
  • A2A: 52,926 queries, 52,552 gallery images
  • A2G: 106,927 queries, 7,959 gallery images
  • G2A: 7,959 queries, 106,927 gallery images
# Method mAP Rank-1 Rank-5 Rank-10 Paper/Code Submitted
1 PersonViT Transformer 9.9% 8.8% 14.4% 17.6% Paper 2025-04-20
2 SeCap Vision-Language 11.2% 8.2% 13.0% 16.2% Paper Code 2025-04-18
3 CLIP-ReID Vision-Language 9.5% 8.9% 12.8% 15.3% Paper Code 2025-04-10
# Method mAP Rank-1 Rank-5 Rank-10 Paper/Code Submitted
1 PersonViT Transformer 22.3% 19.6% 24.8% 27.6% Paper 2025-04-20
2 CLIP-ReID Vision-Language 22.0% 19.7% 24.0% 26.2% Paper Code 2025-04-10
3 SeCap Vision-Language 20.5% 18.1% 21.5% 23.4% Paper Code 2025-04-18
# Method mAP Rank-1 Rank-5 Rank-10 Paper/Code Submitted
1 CLIP-ReID Vision-Language 20.8% 58.1% 63.1% 65.2% Paper Code 2025-04-10
2 PersonViT Transformer 23.3% 51.9% 59.4% 63.0% Paper 2025-04-20
3 SeCap Vision-Language 21.2% 50.9% 57.7% 60.7% Paper Code 2025-04-18

Multi-View Tracking

Task Description

The multi-view tracking task evaluates a model's ability to maintain identity consistency across multiple drone viewpoints, including varying altitudes, angles, and environmental conditions.

Metrics
  • MOTA: Multi-Object Tracking Accuracy
  • IDF1: ID F1 Score (identity preservation)
  • HOTA: Higher Order Tracking Accuracy
  • FP: False Positives
  • FN: False Negatives
  • IDs: ID Switches
Evaluation Protocol

Models are evaluated on their ability to track multiple identities across two challenging scenarios:

  • Single-View: Continuous tracking across a single drone trajectory
  • Multi-View: Identity association across multiple drone viewpoints
# Method HOTA ↑ MOTA ↑ IDF1 ↑ FP ↓ FN ↓ IDs ↓ Paper/Code Submitted

Action Recognition

Task Description

The action recognition task evaluates a model's ability to identify human activities from aerial viewpoints, often with limited resolution and challenging viewpoints.

Action Classes

DetReIDX includes 13 action classes:

  • Walking
  • Running
  • Standing
  • Sitting
  • Cycling
  • Exercising
  • Petting
  • Talking on Phone
  • Leaving Bag
  • Fall
  • Fighting
  • Dating
  • Offending/Trading
Metrics
  • Accuracy: Top-1 classification accuracy
  • mAP: Mean Average Precision across all classes
  • F1: Harmonic mean of precision and recall
# Method Accuracy ↑ mAP ↑ F1 ↑ D1 (≤20m) ↑ D2 (20-50m) ↑ D3 (≥50m) ↑ Paper/Code Submitted

Performance Visualizations

Interactive visualizations of model performance across different tasks and metrics.

Detection Performance by Method
Performance by Distance
Performance by Angle
Re-ID Performance Comparison
A2A vs A2G vs G2A Performance
Rank-1/5/10 Analysis
Performance Degradation by Distance
Model Robustness Analysis
Height vs. Distance Impact