Reproduce “CORE-ReID V2: Advancing the Domain Adaptation for Object Re-Identification with Optimized Training and Ensemble Fusion” on OpenResearch

About this paper

Trinh Quoc Nguyen, Oky Dicky Ardiansyah Prima, Syahid Al Irfan, Hindriyanto Dwi Purnomo, Radius Tanone

CORE-ReID V2 advances Unsupervised Domain Adaptation for object re-identification by introducing an optimized framework that effectively handles both person and vehicle re-identification. It achieves state-of-the-art or competitive performance across various benchmarks, including improvements of up to 16.1% mAP on Person ReID tasks and 5.21% mAP on Vehicle ReID tasks, while supporting lightweight backbone architectures for enhanced efficiency.

Key insights

• Adaptive attention mechanisms that enhance both local (via ECAB) and global (via SECAB) features are crucial for learning more discriminative representations and improving UDA performance across varied object ReID tasks.
• Robust clustering initialization, specifically with Greedy K-means++, consistently leads to higher quality pseudo-labels, thereby enhancing the stability and accuracy of the UDA training process compared to random initialization.
• The ability to generalize ReID frameworks to lightweight backbone architectures without significant performance degradation is vital for practical deployment in resource-constrained environments, offering a favorable trade-off between accuracy and computational efficiency.

Summary by alphaXiv.