The multiplication of publicly available datasets makes it possible to develop Deep Learning models for many real-world applications. However, some domains are still poorly explored, and their related datasets are often small or inconsistent. In addition, some biases linked to the dataset construction or labeling may give the impression that a model is particularly efficient. Therefore, evaluating a model requires a clear understanding of the database. Moreover, a model often reflects a given dataset's performance and may deteriorate if a shift exists between the training dataset and real-world data. In this paper, we derive a more consistent and balanced version of the TrashCan [6] image dataset, called UNO, to evaluate models for detecting non-natural objects in the underwater environment. We propose a method to balance the number of annotations and images for cross-evaluation. We then compare the performance of a SOTA object detection model when using TrashCAN and UNO datasets. Additionally, we assess covariate shift by testing the model on an image dataset for real-world application. Experimental results show significantly better and more consistent performance using the UNO dataset.