Reproduce “Centroid-UNet: Detecting Centroids in Aerial Images” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
N. Lakmal Deshapriya, Dan Tran, Sriram Reddy, Kavinda Gunasekara
CENTROID-UNET adapts the U-Net architecture for detecting object centroids in aerial images by using fixed-size Gaussian kernels for ground truth representation. This method streamlines data annotation while achieving F1-scores of 0.758 for buildings and 0.839 for coconut trees, demonstrating a practical approach for large-scale remote sensing applications.
Key insights
- Adapting a standard semantic segmentation network like U-Net with a straightforward Gaussian kernel representation for centroids can effectively localize objects.
- Reducing the annotation burden to only object centroids is sufficient for many large-scale remote sensing tasks, avoiding the need for precise object boundaries.
- The approach simplifies the deep learning pipeline for object localization, making it more accessible and computationally lighter than methods requiring adaptive kernels or complex loss functions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.