Figure 3.

StackBox over the test data. In Step 1, the predictions of each base learner are considered as the ground truth, one at a time, and the predictions returned by the remaining models that have the highest IoU with the ground truth are chosen. In Step 2, similar to the method used in the training data, the ground truths without associated predictions are removed. When the number of predictions is lower than the number of base learner models used minus one, the null values are filled with the prediction where the IoU with the ground truth is higher. In Step 3, all matches obtained in the previous step are concatenated, and the duplicates are removed. In Step 4, four data sets are created according to the coordinates available for each bounding box. The predictors include the corresponding coordinate of each base learner, and the target is predicted using the meta learner models obtained in training data (Step 5). In Step 6, the predictions are combined, and NMS (with a threshold of 0.5) is applied to remove redundant boxes. The red boxes represent the predictions of the base learner models in the test data.