GIC presented novel research for detecting centroids of objects in aerial images at the 42nd Asian Conference on Remote Sensing (ACRS) which was held from November 22-24, 2021 in Can Tho, Vietnam.

GIC Team Leader Lakmal  Deshapriya delivered a presentation describing CentroidUNet, a new approach to detecting centroids
of objects appearing in RGB aerial and satellite imagery with a deep neural network.

The Centroid-UNet model is based on classic U-Net semantic segmentation architecture, which has been optimized for centroid detection. The technique uses Gaussian blobs generated around centroid points to stabilize the training process. The model was tested and evaluated in a case study using high resolution aerial
imagery for centroids of buildings and coconut trees. The simplified approach achieved a good accuracy for both object types in comparison to other centroid detection methods.

The next steps will assess the feasibility of using this approach for multi-class centroid estimation. The code for Lakmal’s Centroid-UNet research as well as the trained models are published in GitHub –

Due to the ongoing COVID-19 pandemic, the majority of the 42nd meeting of ACRS was held online.

ACRS is an annual conference hosted by member countries of the Asian Association on  Remote Sensing (AARS). The conference dates back to 1980 with an inaugural meeting in Bangkok, Thailand. Professor Shunji Murai, organizer and catalyst for the first ACRS conference, continues to be an integral part of its success to this day.

In 2022 ACRS will be holding its 43rd meeting in Ulaanbaatar, Mongolia. Stay tuned to the AARS website for future ACRS announcements.