Small object detection in image with transfer learning

Abstract

Small object detection is a branch of computer vision and it has many applications for our daily life such as monitoring security, traffic tracking, etc. In this thesis, small objects are in the context of input images from surveillance cameras; so the objects are not necessary to be small, but they appear small in input images. Small objects in digital images are presented with only few pixels, which makes them challenging to detect due to lack of information. With rapid advancement of deep learning, small object detection can be tackled more effectively than ever. In this thesis, some novel convolutional neural networks, namely Faster R-CNN and Retina-Net, were experimented with different setups of backbone networks to see how they handle small object detection task. The networks were trained with transfer learning technique on VisDrone2018 training dataset and then the inferences were validated on validation test for analysis. The experimental results showed that Faster R-CNN network with ResNext101 and FPN as backbone produces the highest mAP score of 20.950% from all selected setups. However, this mAP score is still far behind HAL-Retina-Net which is the winner of object detection in image track in VisDrone2018 Challenge with 31.88% mAP score. After the experiments, it was concluded that Faster R-CNN and RetinaNet alone are not enough for small object detection. Additionally, many difficulties were discovered, and some proposals were dis-cussed to improve the task in the future.