Accurate whole heart segmentation can enhance the modeling and analysis of cardiovascular disease treatment. However, real-world imaging analysis faces challenges in generalization due to data inhomogeneity, which arises from factors such as low-quality image acquisition, variations across multiple modalities, differences in scanner vendors, and cardiac motion. In this paper, to tackle the problem of data inhomogeneity, we use data augmentation and model calibration for whole heart segmentation enhancement. We first adopted the state-of-the-art MedNeXt transformer-based architecture as the baseline. Then, we enhance the baseline in model robustness through an ensemble strategy with strong data augmentation and out-of-distribution model calibration techniques. Comprehensive experiments have been conducted on the dataset from the CARE2024 Task 5 Whole Heart Segmentation (WHS++) challenge, in which 7 target structures from CT and MRI images acquired from multiple centers are considered. Results show that data augmentation and model calibration can effectively improve the segmentation performance across various modalities and centers. We highlight that our team ranks first on the validation leaderboard with average dice scores of 0.9440 (CT) and 0.8956 (MRI).