Classification and localization of extreme outliers in computer vision tasks in surveillance scenarios

Sammanfattning: Convolutional neural networks (CNN) have come a long way and can be trained toclassify many of the objects around us. Despite this, researchers do not fullyunderstand how CNN models learn features (edges, shapes, contours, etc.) fromdata. For this reason, it is reasonable to investigate if a CNN model can learn toclassify objects under extreme conditions. An example of such an extreme conditioncould be a car that drives towards the camera at night, and therefore does not haveany distinct features because the light from the headlights covers large parts of thecar.The aim of this thesis is to investigate how the performance of a CNN model isaffected, when trained on objects under extreme conditions. A YOLOv4 model willbe trained on three different extreme cases: light polluted vehicles, nighttimeobjects and snow-covered vehicles. A validation will then be conducted on a testdataset to see if the performance decreases or improves, compared to when themodel trained is on normal conditions. Generally, the training was stable for allextreme cases and the results show an improved or similar performance incomparison to the normal cases. This indicates that models can be trained with allextreme cases. Snow-covered vehicles with mosaic data augmentation and the IOUthreshold 0,25 had the best overall performance compared to the normal cases, witha difference of +14,95% in AP for cars, −0,73% in AP for persons, +8,08% in AP fortrucks, 0 in precision and +9% in recall.