Assistive robots should be able to wash, fold or iron clothes. However, due to the variety, deformability and self-occlusions of clothes, creating general-purpose robot systems for cloth manipulation is challenging. Synthetic data is a promising direction to improve generalization, though its usability is often limited by the sim-to-real gap. To advance the use of synthetic data for cloth manipulation and to enable tasks such as robotic folding, we present a synthetic data pipeline to train keypoint detectors for almost flattened cloth items. To test its performance, we have also collected a real-world dataset. We train detectors for both T-shirts, towels and shorts and obtain an average precision of 64.3%. Fine-tuning on real-world data improves performance to 74.2%. Additional insight is provided by discussing various failure modes of the keypoint detectors and by comparing different approaches to obtain cloth meshes and materials. We also quantify the remaining sim-to-real gap and argue that further improvements to the fidelity of cloth assets will be required to further reduce this gap. The code, dataset and trained models are available online.