Figure 8
From: Aphid cluster recognition and detection in the wild using deep learning models
The IoU threshold in NMS algorithm v.s. average precision. Similar to Table 2, “original” indicates the dataset is our originally labeled dataset; “10 pixels” represents the neighboring clusters within 10 pixels are merged with a single bounding box; “10 pixels + rm 0.01” illustrates that merging neighboring clusters within 10 pixels and removing tiny clusters (less than 0.01 of the area of the patches) are implemented for the dataset. Split 1 is utilized as the testing data and the other 9 splits are merged as the training data. The four detection models have the same trend and the performance is extremely unpleasing when the threshold is close to 1. The IoU threshold of around 0.5 could yield the best performance with all detection models and annotations. In the experiments, 0.6 is utilized as the default IoU threshold for NMS algorithm and all the experimental results are based on 0.6 as the IoU threshold of NMS.
