Figure 5
From: Aphid cluster recognition and detection in the wild using deep learning models
Several examples of image patches contain extremely small partial ground truth after cropping. In most cases, the small partial ground truth bounding boxes appear at the border of the image patches, which is illustrated above with red bounding boxes. If those image patches are in the training set, the contribution of those patches is almost zero for training the detection models. If those image patches are in the testing set, the overall performance would drop since even the best detection model cannot accurately localize those small partial aphid clusters at the border. Thus removing those small partial aphid clusters is meaningful in both training and evaluation.
