Object detection in autonomous driving - from large to small datasets

Abstract

The purpose of the paper is to analyze the current capacity of pedestrian and vehicle detection through four state of the art detectors -Yolo, SSD, Faster R-CNN and RetinaNet on a big dataset (BDD100K). Also, we analyzed if the results are transferable from one dataset to another - we used a small dataset from our campus, we offered some quantitative results and we made an error analysis based on the dataset characteristics (e.g. weather, light, size of the object).