This paper introduces a novel approach to unsupervised object Re-Identification (ReID) by focusing on feature calibration, a conceptually simple yet empirically powerful method. The authors propose a Feature Calibration Module (FCM) that operates before pseudo-label generation, using a nonparametric graph attention network to refine the extracted features. This module ensures that similar instances cluster together and dissimilar instances separate, improving the quality of pseudo-labels and subsequent representation learning. The FCM is parameter-free, training-free, and can be easily integrated into existing clustering-based ReID methods without significant impact on training or testing efficiency. Extensive experiments on various benchmarks show that the proposed method consistently outperforms baseline methods, achieving state-of-the-art results, particularly on challenging datasets like MSMT17. The authors also provide in-depth analysis to validate the effectiveness of the FCM, demonstrating its ability to enhance clustering quality and improve overall performance.This paper introduces a novel approach to unsupervised object Re-Identification (ReID) by focusing on feature calibration, a conceptually simple yet empirically powerful method. The authors propose a Feature Calibration Module (FCM) that operates before pseudo-label generation, using a nonparametric graph attention network to refine the extracted features. This module ensures that similar instances cluster together and dissimilar instances separate, improving the quality of pseudo-labels and subsequent representation learning. The FCM is parameter-free, training-free, and can be easily integrated into existing clustering-based ReID methods without significant impact on training or testing efficiency. Extensive experiments on various benchmarks show that the proposed method consistently outperforms baseline methods, achieving state-of-the-art results, particularly on challenging datasets like MSMT17. The authors also provide in-depth analysis to validate the effectiveness of the FCM, demonstrating its ability to enhance clustering quality and improve overall performance.