inspection. Cows detected correctly but with a wrongly predicted id at the feeding table are given by the “ wrongly predicted cow-id” category. The “ wrongly detected cows” contains cows where the detection of the cow in the image failed which makes the cow-id prediction processes impossible. Table 1. Summary statistics for comparison of the contour based prediction algorithm with manual labels . Sample Correctly predicted cow-id Wrongly predicted cow-id Wrongly detected cows Count 5711 335 311 Fraction [%] 90 5 5 The pairwise error is defined as the number of times one cow is misclassified as another cow normalized by number of observations of the first cow. It can be used to evaluate the uniformity of the errors between the different classes. The distribution of the pairwise errors lager than zero for all cows in the validation set is given in Figure 2A. The F1-score is used to evaluate the performance of the classifier for each cow. The F1-score measures both the ability of the model to predict the cow-id correctly and the ability not to give this id to a cow with another id. The distribution over the F1-scores for cows can be seen in Figure 2B. The F1-scores does not include the error rates described by the last category in Table 1 and should therefore only be regarded as a relative measure of the accuracy for the different individuals. Figure 2. A, histogram of the non-zero pairwise error rates observed on the validation data. (Note the logarithmic y-axis.) B, histogram showing the distribution of F1-scores (one vs. all) for individual cow. The F1-scores does not include the error rates described by the last three categories in Table 1. Discussion With a fraction of 90% correctly predicted cow-id this study clearly demonstrates that the geometry of the cow back region is unique to each individual. A similar relation is well known for other species, for instance the human facial region (Bowyer et al., 2006). The distribution of the F1-scores for each of the 97 cows in the validation experiment can be partitioned into two as seen in Figure 2 B. A high accuracy group with F1-score above 0.87 and a tail group with cows where the prediction algorithm is less accurate and F1-score below 0.87. The latter group of cows are to some extent characterized by having much higher pair wise error rates than the rest of the cows. For instance, the 5 cows with a pair wise error above 10 % are all present in this region. These cows are in other words very similar to other cows in the validation set and as a result the classifier is less accurate for these. However, the
Download PDF file