Achieving a high-quality and fine-grained appearance representation is an important task in Multi-Object Tracking (MOT). Recent works learn responses of different locations or channels from the global feature map to enhance visible parts or weaken undesirable parts. However, only relying on global learning actually weakens the gradient of local feature response changes, resulting in loss of details. Thus, the imbalance between global and local learning becomes a key issue in fine-grained representation in MOT. To address this problem, this paper proposes a Multi-Path Enhancer (MPE) to balance global and local extraction.We firstly propose a Multi-Path Part Enhancer (MPPE) to enhance part features. MPPE constructs multiple pathways for parts and adopt a shared parameter module to learn feature responses for each part, achieving efficient local feature enhancement. Secondly, we propose a Global Enhancer (GE). We aggregate the output part responses from MPPE to be the global response and act on the global feature map. Thus GE keeps a global learning pathway without introducing any learnable parameter. We then get a Multi-Path Enhancer by parallelizing the MPPE and GE. Furthermore, we propose a finer-grained training data augmentation method Gaussian Erasing Augmentation (GEA). Based on erasing augmentation, GEA adopts a 2D anisotropic Gaussian kernel to make erasing mask with free-form shapes, which provides more robust training data. Finally, we choose YOLOX as the detector and apply MPE in the appearance representation network, as well as exploit GEA in training. Then we form a clean tracker, MPETrack. Without introducing any extra training data, it gets 64.5 HOTA on MOT17 and 63.1 HOTA on MOT20 challenges, achieving state-of-the-art tracking performance.