umair894/KAT-ReID-Veri776

KAT-ReID – VeRi-776

Model Details

• Model name: KAT-ReID (VeRi-776)
• Architecture: Kolmogorov–Arnold Transformer (KAT) with GR-KAN channel mixers
• Task: Vehicle Re-Identification
• Dataset: VeRi-776
• Framework: PyTorch
• License: MIT
• Paper: KAT-ReID: Assessing the Viability of Kolmogorov–Arnold Transformers in Object Re-Identification

This model replaces the MLP blocks of a ViT-based ReID backbone with Group-Rational Kolmogorov–Arnold Networks (GR-KAN) while retaining self-attention and ReID-specific architectural components.

---

Model Description

The model is trained for vehicle re-identification, where the goal is to retrieve images of the same vehicle across different cameras and viewpoints.

Key architectural features:

• GR-KAN replaces standard MLP channel mixers
• Side-information embedding (camera/view conditioning)
• Local token rearrangement branch to preserve spatial cues
• Joint optimization with ID classification and metric learning losses

---

Training Data

• Dataset: VeRi-776
• Identities: 776 vehicles
• Images: 49,357
• Cameras: 20
• Views: 8

Training follows the official dataset split and evaluation protocol.

---

Training Procedure

• Pretraining: ImageNet-1K
• Input resolution: 256 × 128
• Patch size: 16 × 16 (overlapping stride 12)
• Optimizer: SGD (lr=0.008, momentum=0.9)
• Batch size: 64 (16 IDs × 4 images)
• Losses: Cross-Entropy (ID) + Triplet loss
• Augmentations: Random flip, random erasing
• Mixed precision: Enabled

---

Evaluation Results

Metric	Score
mAP	59.5
Rank-1	88.0
Rank-5	95.8
Rank-10	98.0

Results are reported under single-query evaluation without re-ranking.

---

Intended Use

This model is intended for:

• Academic research in ReID
• Benchmarking alternative transformer channel mixers
• Studying robustness under viewpoint variation

Not intended for real-world surveillance or deployment without further validation.

---

Limitations

• Underperforms strong ViT baselines on globally discriminative vehicle benchmarks
• Sensitive to training stability due to rational activation parameterization
• Performance may vary with different camera distributions

---

Citation

@inproceedings{umair2025katreid,
  title={KAT-ReID: Assessing the Viability of Kolmogorov--Arnold Transformers in Object Re-Identification},
  author={Umair, Muhammad and Zhou, Jun and Musaddiq, Muhammad Hammad and Muhammad, Ahmad},
  year={2025}
}