The task of automatic visual inspection has been tackled by numerous machine learning algorithms. However, there are no global comparative studies of the performance of these algorithms that use metrics directly relevant to their use in an industrial setting. We survey the performance of machine learning algorithms applied to the DAGM dataset, a reference dataset for industrial visual inspection. Our study reports the performance of 17 algorithms for which the learning and evaluation protocols are clear enough to be reproducible. However, not all of these algorithms are comparable, in the sense that they rely on different labelling or even different data. We group the comparable results, and conclude that the DAGM dataset no longer presents a major difficulty for the algorithms based on knowledge of the location of defects in the images. On the other hand, algorithms using only unlabelled images, which are the easiest to implement in practice, do not yet achieve industrially acceptable performance.