Department of Computer Engineering
Object Detection in Remote Sensing Imagery
(Supervisor: Assoc. Prof. Dr. Selim Aksoy)
Computer Engineering Department
Many efforts have been devoted recently into various methods for object detection in remote sensing images. However, the progress in the field is not as advanced as natural scene object detection. Natural scene images contain information of objects sidewards and in profile, whereas the remote sensing images capture the objects from a bird's-eye view. Therefore, the object detectors learned from natural scene datasets are not easily applicable to aerial images. Also, most of the existing remote sensing object detection datasets have some limitations and the number of datasets in the domain is limited. Aerial datasets are not as inclusive as the ones in natural scene object detection (ImageNet, MS COCO). In our research, we work on applying deep learning techniques to detect objects in remote sensing images. Two-stage and one-stage object detection methods are applied in DOTA which is a relatively recent released large-scale dataset. There are speed and accuracy trade-offs between these convolutional object detectors. More specifically, two-stage object detectors use a Region Proposal Network to generate some region proposals in the first stage and feed the region proposals into the object classification and bounding-box regression in the second stage. On the other hand, one-stage object detectors use predefined candidate object locations regularly sampled across an image and learn the class probabilities and bounding box coordinates from the input image. Compared to two-stage counterparts, one-stage object detectors are much faster however they have low accuracy rates. In this presentation, we discuss the two main types of object detectors, application of object detector models on DOTA dataset and compare it with the current baselines.
DATE: 25 November 2020, Wednesday @ 13:35