The process of labeling data for intricate and specialized downstream tasks, including 2D human pose estimation, necessitates extensive expertise and can be financially burdensome. Consequently, pseudo-labels are increasingly becoming favored alternatives. In the context of 2D human pose estimation, frequent inter or intra-class imbalances often exhibit diverse locating challenges for poses due to variations in scale, motion, and occlusion. Such imbalances can potentially cause data bias and initiate self-training bias, ultimately resulting in a model collapse and increased model errors. Despite previous research efforts that have aimed to rectify model collapses through augmentation techniques and enhance model robustness via dual network training, two substantial issues remain unaddressed: (1) the learned representations often suffer from imbalanced distribution, and (2) this imbalance results in an increased amount of incorrect pseudo-labeling. To address these critical limitations, this paper introduces a novel reverse re-balancing module. This module optimizes representations to solve data bias caused by upstream imbalances. To underscore the module’s efficacy, we have also proposed a data augmentation method tailored to it, identified as adaptive joint CutPaste. Furthermore, to mitigate the second highlighted problem, we propose deploying co-guessing techniques to amalgamate the pseudo labels. The effectiveness of our newly developed approach, Debiasing-Teacher, is scrutinized using two benchmark datasets designed explicitly for pose estimation—MSCOCO and MPII. Compared to preceding methods, our approach shows significant improvements, particularly in limited annotation scenarios. The source codes for Debiasing-Teacher can be accessed at https://github.com/wangnaihao/Debias-Teacher .