Search

The a priori probability of a dataset is usually used as a baseline for comparing a particular algorithm’s accuracy in a given binary classification task. ZeroR is the simplest algorithm for this, predicting the majority class for all examples. However, this is an extremely simple approach that has no predictive power and does not describe other dataset features that could lead to a more demanding baseline. In this paper, we present the Extended A Priori Probability (EAPP), a novel semi-supervised baseline metric for binary classification tasks that considers not only the a priori probability but also some possible bias present in the dataset as well as other features that could provide a relatively trivial separability of the target classes. The approach is based on the area under the ROC curve (AUC ROC), known to be quite insensitive to class imbalance. The procedure involves multiobjective feature extraction and a clustering stage in the input space with autoencoders and a subsequent combinatory weighted assignation from clusters to classes depending on the distance to nearest clusters for each class. Class labels are then assigned to establish the combination that maximizes AUC ROC for each number of clusters considered. To avoid overfit in the combined feature extraction and clustering method, a cross-validation scheme is performed in each case. EAPP is defined for different numbers of clusters, starting from the inverse of the minority class proportion, which is useful for a fair comparison among diversely imbalanced datasets. A high EAPP usually relates to an easy binary classification task, but it also may be due to a significant coarse-grained bias in the dataset, when the task is previously known to be difficult. This metric represents a baseline beyond the a priori probability to assess the actual capabilities of binary classification models.

Chest X-ray images are useful for early COVID-19 diagnosis with the advantage that X-ray devices are already available in health centers and images are obtained immediately. Some datasets containing X-ray images with cases (pneumonia or COVID-19) and controls have been made available to develop machine-learning-based methods to aid in diagnosing the disease. However, these datasets are mainly composed of different sources coming from pre-COVID-19 datasets and COVID-19 datasets. Particularly, we have detected a significant bias in some of the released datasets used to train and test diagnostic systems, which might imply that the results published are optimistic and may overestimate the actual predictive capacity of the techniques proposed. In this article, we analyze the existing bias in some commonly used datasets and propose a series of preliminary steps to carry out before the classic machine learning pipeline in order to detect possible biases, to avoid them if possible and to report results that are more representative of the actual predictive power of the methods under analysis.

Pedestrian detection is a particular case of object detection that helps to reduce accidents in advanced driver-assistance systems and autonomous vehicles. It is not an easy task because of the variability of the objects and the time constraints. A performance comparison of object detection methods, including both GPU and non-GPU implementations over a variety of on-road specific databases, is provided. Computer vision multi-class object detection can be integrated on sensor fusion modules where recall is preferred over precision. For this reason, ad hoc training with a single class for pedestrians has been performed and we achieved a significant increase in recall. Experiments have been carried out on several architectures and a special effort has been devoted to achieve a feasible computational time for a real-time system. Finally, an analysis of the input image size allows to fine-tune the model and get better results with practical costs.

Vulnerable Road User (VRU) detection is a major application of object detection with the aim of helping reduce accidents in advanced driver-assistance systems and enabling the development of autonomous vehicles. Due to intrinsic complexity present in computer vision and to limitations in processing capacity and bandwidth, this task has not been completely solved nowadays. For these reasons, the well established YOLOv3 net and the new YOLOv4 one are assessed by training them on a huge, recent on-road image dataset (BDD100K), both for VRU and full on-road classes, with a great improvement in terms of detection quality when compared to their MS-COCO-trained generic correspondent models from the authors but with negligible costs in forward pass time. Additionally, some models were retrained when replacing the original Leaky ReLU convolutional activation functions from original YOLO implementation with two cutting-edge activation functions: the self-regularized non-monotonic function (MISH) and its self-gated counterpart (SWISH), with significant improvements with respect to the original activation function detection performance. Additionally, some trials were carried out including recent data augmentation techniques (mosaic and cutmix) and some grid size configurations, with cumulative improvements over the previous results, comprising different performance-throughput trade-offs.