The dataset is crucial for the results of crack segmentation in deep learning. However, the quantity and quality of annotations in datasets used for crack segmentation are uneven, and there is a lack of dataset benchmarks. To address this issue, this paper proposes a dataset benchmark for pixel-level segmentation of concrete cracks based on a deep learning model. Firstly, the ability of Resnet-101, DRN, Xception, and Mobilenetv2 to extract crack features is evaluated to preferably select an appropriate backbone network for the DeepLabv3 + model. Secondly, based on the deep learning model and crack dataset, a universal calculation model is proposed to estimate the quantity benchmark of high-quality crack annotated samples, as well as the mixture ratio of high-quality and low-quality annotated samples. Additionally, a pixel-level visualization characterization method is constructed for inference results of crack edge details. The results show that the model with Resnet-101 as the backbone network exhibits significant performance, with pixel accuracy (PA), mean intersection-over-union (mIoU), and frequency-weighted intersection-over-union (FWIoU) reaching 99.76%, 95.05%, and 99.52%, respectively. The combination ratio of low-quality and high-quality annotated crack images has different effects on the model performance for datasets with different number of samples. The proposed benchmark for the concrete crack dataset holds engineering value and significance, and can be applied to improve model performance by accurately selecting appropriate methods, such as increasing the number of samples in the dataset or optimizing the model architecture, while minimizing the annotation time of the dataset.