In this video, I will talk a little bit about detector training. How many objects and background regions are in the image? The number of object is several magnitudes times smaller than the number of background regions. So, object detection is an essentially asymmetrical problem with very imbalanced classes. The class background is very difficult, so we need a lot of good data for training. And good training requires similar number of objects and backgrounds examples. I will demonstrate the training procedure using the examples from Andrew Zisserman VGG group. That's the training of histogram of orientation detector plus SVM for person upper torso. The data set is created from 33 videos from Hollywood films. Totally, 1122 frames were annotated, with more than 1600 persons in total. First, positive data is augmented. Next after data augmentation for object detector training are the same as for image classification training. We apply random augmentations like scaling, shifts, rotation, mirroring to positive examples. In this particular example, from 1600 ground truth examples, we get 32,000 of jittered examples. The only way to populate the negative data set, is to use random negative examples. As seen in this slide, most of such examples should be very easy for the detector. Then, we train first detector using augmented positive data and random negative data. The first detector is applied to train images. Our goal is to search for high-confidence false detections. Such detections are hard examples for the detector because it fails on them. We then add this hard-negative examples to the data set. The classifier is then re-trained in the procedure of hard-negative mining is repeated. Here, you can see examples for false detections. Most of them are close to true detections but with very bad localization. After several alteration, we plugged the precision recall curve of the final detector against the first detector. You can clearly see that the detector performance has improved dramatically. So, hard examples are very important for good detector training. What if negative data set is constructed only from hard negative examples? It can degrade the performance. So we usually need both easy and hard examples for training of the detector. On this slide, the detection from first and final detectors are demonstrated for two images. You can see that the number of false detection has lowered. But on the second image, one of the true detection is missing. Overall, hard-negative mining procedure can greatly improve the performance of the detector. But we should be sure that all objects has been annotated. Otherwise, missed objects can be also detected and added to the negative examples, and this will greatly degrade the performance of the detector.