Enhancing Target Re-identification via Model Fusion and Knowledge Distillation of Pre-trained Foundation Models

Tendai Shoko, Terence L. van Zyl

Abstract: Target re-identification (re-ID) systems face critical deployment challenges balancing accuracy with computational efficiency in resource-constrained environments. This paper presents a novel framework that integrates Mixture-of-Experts (MoE) with Knowledge Distillation (KD) to effectively leverage pretrained foundation models. The framework employs dynamic expert selection to combine CLIP and ALIGN models, then distils their collective knowledge into a compact student architecture. The experimental evaluation on VeRi-776 and Market-1501 demonstrates 75. 2% and 76. mAP 1%, respectively, while reducing the inference time by 50% and the model parameters by 94% compared to the MoE ensemble (and approximately 92% vs. CLIP fine-tuning). Comprehensive ablation studies validate the synergistic benefits of MoE and KD components, showing improved cross-domain performance with 12.9% mAP degradation versus 15.3% for conventional methods. The results demonstrate MoE-KD as a practical solution for real-world reID deployment.

External IDs:doi:10.1007/978-3-032-11733-5_23