This paper explores the topic of anonymous open-world cyclist re-identification. Person re-identification (re-ID) with deep neural networks has made progress and achieved high performance in recent years. However, most existing re-ID works are designed for closed-world scenarios rather than realistic open-world settings, which limits the practical application of the re-ID technique. Currently, no dataset of cyclists exists. Directly applying a trained re-ID network on another dataset does not yield good results. Therefore, a new dataset of cyclists is introduced in this paper. Our dataset is different than most existing benchmark datasets as every person in our dataset has been blurred to respect their privacy. In this paper the effect of blurring on re-identification performance is evaluated. To evaluate the impact of blurring on the re-identification performance we first tested it on the Market1501 dataset. Here, the performance of the blurred version could easily be compared to the original version blurring. The experiments show that blurring the data only impacts the rank-1, and mAP score by 1-4% for the Market1501 dataset. This impact depends on the size of the blurring window that is used. Several state of the art performing re-identification models were rebuilt and evaluated on our new dataset and their performance was compared. Furthermore, different backbone architectures were evaluated, we found that EfficientNetB0 outperforms the standard ResNet50 backbone architecture for re-identification, while using fewer parameters. Next the effects of RandAugment and Cosine learning rate decay were evaluated for re-identification. It was found that including RandAugment increases the rank-1 and mAP scores achieved on our dataset by up to 3%, and that using cosine decay further improves the achieved score. The final scores achieved on our dataset are 89.8% rank-1 accuracy and a mAP of 81.4%. Next we show that the batch hard pairwise loss function increases the F1-score by 7% for open-world re-identification. It was concluded that combining the embeddings is necessary to achieve good performance for open-world re-identification.