Prediction vs label performance of the network on the test set consisting of 347 objects. On the x-axis, we show the 21 soft labels from our visual classification. On the y-axis, we show the mean predictions of the model in the 21 soft-label bins. The data point size reflects the number of examples in each bin, with the smallest points (e.g., two points between 0.1 and 0.2), representing bins containing only a single data point. Error bars represent the standard deviation of predictions within each bin. We assumed a Gaussian distribution to compute the mean and standard deviation. The red dashed line shows the one-to-one relation, indicating ideal calibration.