umair894/KAT-ReID-MSMT17

KAT-ReID – MSMT17

Model Details

• Model name: KAT-ReID (MSMT17)
• Task: Large-scale Person Re-Identification
• Dataset: MSMT17

---

Model Description

MSMT17 is a large-scale, multi-camera ReID benchmark with strong domain shifts. This model evaluates KAT-ReID under challenging illumination and camera variability.

---

Training Data

• Identities: 4,101
• Images: 126,441
• Cameras: 15

---

Evaluation Results

Metric	Score
mAP	38.8
Rank-1	61.5
Rank-5	76.6
Rank-10	82.2

Results show that KAT-ReID struggles to match ViT baselines on highly diverse large-scale datasets.

---

Limitations

• Underfits global appearance cues
• Requires further regularization and longer warm-up
• Sensitive to large domain shifts

---

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}
}