Human Identification at a Distance: Challenges, Methods and Results on the Competition HID 2025

Technical Summary: Human Identification at a Distance (HID) 2025 Competition

1. Problem Statement

Human Identification at a Distance (HID) is a critical challenge in public security and surveillance. Traditional biometrics, such as facial recognition and fingerprinting, often fail in long-range monitoring or crowded environments where high-resolution imagery is unavailable. Gait recognition—identifying individuals by their unique walking patterns—is a viable alternative because it can be captured from a distance using silhouettes.

However, gait recognition faces significant hurdles in "in-the-wild" scenarios, including:

• Appearance variations: Changes in clothing and carried objects.
• Viewpoint variations: Different camera angles.
• Domain shifts: The gap between controlled laboratory training data and unconstrained real-world deployment data.

2. Methodology (Competition Setup)

The HID 2025 competition was designed to push the boundaries of gait recognition robustness.

• Dataset: It utilized the SUSTech-Competition dataset, characterized by significant variations in clothing, carried items, and camera viewpoints. To prevent overfitting, a new random split was used for the 2025 test set.
• Challenge Constraints: No training data was provided. Participants had to leverage external datasets (e.g., CASIA-B, Gait3D, etc.) to train their models, creating a strict cross-domain generalization requirement.
• Evaluation: The primary metric was Rank-1 Accuracy. The competition employed a two-phase evaluation (a 10% subset followed by the full test set) to mitigate label leakage and ensure reproducibility.

3. Key Technical Trends & Top-Performing Methods

The paper analyzes the methodologies of the top nine teams, revealing several dominant technical trends:

• Backbone Architectures: There was a strong consensus on using DeepGaitV2 (specifically P3D and 3D convolutional modules) as the foundational architecture. Some teams integrated Transformer-based models like SwinGait to capture long-range dependencies.
• Spatial Alignment: To standardize walking postures, most top teams implemented unsupervised alignment (e.g., PCA-based rotation). The winning team (BRAVO-FJ) achieved superior results by using a hybrid silhouette- and skeleton-based alignment approach.
• Domain Adaptation & Fine-tuning: A key strategy for the top performers was fine-tuning strong pre-trained models using a small amount of target-domain data (the HID 2025 gallery set) to bridge the domain gap.
• Data Augmentation & Cleaning: Techniques included "random erasing white" (to simulate occlusions), perspective transformations, and rigorous data cleaning (e.g., Connected Component Analysis) to remove structural noise from silhouettes.
• Inference Optimization: Most teams employed Re-ranking and Ensemble/Voting mechanisms (combining multiple model checkpoints or architectures) to stabilize and boost final accuracy.

4. Results

The competition demonstrated significant algorithmic progress:

• New Benchmark: The highest accuracy achieved was 94.2% (by Team BRAVO-FJ), setting a new record for the SUSTech-Competition dataset.
• Performance Trend: The results show a steady upward trajectory in state-of-the-art accuracy over the six years of the HID competition, even as the datasets became increasingly challenging.
• Effectiveness of Strategies: The paper confirms that combining multi-dataset pre-training with targeted fine-tuning and spatial alignment is highly effective for cross-domain gait recognition.

5. Significance and Future Directions

The paper concludes that while gait recognition has reached high levels of accuracy, the field must move toward more complex, real-world scenarios.

Significance:

• It provides a standardized, fair, and rigorous benchmark for evaluating the generalization capabilities of gait recognition models.
• It identifies the current "state-of-the-art" toolkit (DeepGaitV2 + Alignment + Ensemble).

Future Research Directions:

1. Large Vision Models (LVMs): Leveraging foundation models pre-trained on massive video/image datasets to improve zero-shot or few-shot generalization.
2. Multi-modal Integration: Combining silhouettes with 3D structures, skeletal data, or scene context to handle extreme occlusions and viewpoints.
3. Synthetic Data Generation: Using generative AI (e.g., video diffusion models) to create diverse, privacy-preserving synthetic gait datasets to overcome data scarcity.
4. Comprehensive Metrics: Moving beyond Rank-1 accuracy to include Rank-5 and AUCMC to better evaluate retrieval performance.