umair894/KAT-ReID-Occluded-duke

KAT-ReID – Occluded-Duke

Model Details

• Model name: KAT-ReID (Occluded-Duke)
• Task: Person Re-Identification under Occlusion
• Architecture: KAT with GR-KAN channel mixing
• Dataset: Occluded-DukeMTMC
• Framework: PyTorch

---

Model Description

This model targets occlusion-robust person re-identification. It demonstrates the strongest gains of KAT-ReID over ViT baselines, validating the hypothesis that GR-KAN mixers improve partial-observation generalization.

---

Training Data

• Dataset: Occluded-DukeMTMC
• Identities: 1,404
• Images: 36,441
• Cameras: 8

---

Training Procedure

Same pipeline as described in the paper:

• ImageNet initialization
• Overlapping patch embedding
• Side-information embedding (camera ID)
• Local feature branch (K=4)
• SGD optimizer with joint ID + metric loss

---

Evaluation Results

Metric	Score
mAP	69.6
Rank-1	83.7
Rank-5	91.8
Rank-10	94.5

This model outperforms TransReID (ICCV’21) by a large margin under occlusion-heavy conditions.

---

Intended Use

• Research on occlusion-aware ReID
• Partial-observation robustness analysis
• Benchmarking non-MLP transformer architectures

---

Limitations

• Gains are dataset-dependent
• Requires careful initialization and regularization
• Not optimized for real-time inference

---

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}
}