Keywords: object detection, Convolutional Neural Network (cnn), You Only Look Once (yolo), Faster r-cnn (Region-based Convolutional Neural Networks, ssd

Figure 4.1: FlickrLogos-32 dataset

Confusion matrix gives a lot of information about the relation of the model and the real facts. In a confusion matrix we can see the number of true positives (TP). This is the number of the objects the algorithm classified as positive and they are positive. False positive (FP) is the number of objects the algorithm predicted as positive but they are negatives in the real world. False negative (FN) objects are from the negative class according to our model but they are in the positive class in the real world. Finally, true negative objects have been predicted as negatives and in fact they are negatives, too. We try to minimize the number of FPs and FNs, the objects that have been misclassified. Also, we try to maximize the number of TPs and TNs, the number of correct predictions. I present this in an example in Figure 4.3

4.3.2. Precision and recall

Precision and recall are usually not discussed in isolation, they complete each other. Precision
gives information about the fraction of retrieved relevant instances among the retrieved instances, while recall shows the fraction of retrieved relevant instances among the total amount of relevant instances.

precision: the ratio of the true positives and the retrieved instances

(4.1)

recall: the ratio of true positives and the number of the relevant instances

(4.2)

By changing the threshold of what we want our system to consider as relevant, there will be more or less true positives - and also false positives(!). With this method we can characterize a classifier: we look at how precision and recall change as we change the threshold.

Figure 4.2: Detection-results for our project False and True positive
4.5. Result and training process

Using this criterium, we calculate the precision/recall curve. Then we compute a version of the measured precision/recall curve with precision monotonically, by setting the precision for recall r to the maximum precision obtained for any recall r' > r. Finally, we compute the AP as the area under this curve by numerical integration. No approximation is involved since the curve is piecewise constant and finally, we can calculate mean average precision object detection(mAP), resulting in a mAP value from 0 to 100%