The paper "YOLC: You Only Look Clusters for Tiny Object Detection in Aerial Images" addresses the challenges of detecting small objects in large-scale aerial images, which are characterized by their high resolution, non-uniform object distribution, and limited computational resources. The authors propose YOLC (You Only Look Clusters), an anchor-free object detection framework based on CenterNet, which is designed to efficiently handle these challenges. Key contributions include: 1. **Local Scale Module (LSM)**: This module adaptively searches for cluster regions in aerial images and resizes them to the appropriate scale for accurate detection, improving efficiency and accuracy. 2. **Refined Regression Loss**: The regression loss is modified using Gaussian Wasserstein distance (GWD) to better handle small objects, and a combination of GWD and $L_1$ loss is proposed to balance performance for both small and large objects. 3. **Improved Detection Head**: The detection head is enhanced with deformable convolutions to refine bounding box regression and a disentangled heatmap branch to improve localization accuracy for different object categories. The effectiveness of YOLC is demonstrated through extensive experiments on two aerial image datasets, VisDrone2019 and UAVDT, showing superior performance compared to state-of-the-art methods in terms of accuracy and efficiency. The paper also includes ablation studies to validate the effectiveness of each component of the proposed framework.The paper "YOLC: You Only Look Clusters for Tiny Object Detection in Aerial Images" addresses the challenges of detecting small objects in large-scale aerial images, which are characterized by their high resolution, non-uniform object distribution, and limited computational resources. The authors propose YOLC (You Only Look Clusters), an anchor-free object detection framework based on CenterNet, which is designed to efficiently handle these challenges. Key contributions include: 1. **Local Scale Module (LSM)**: This module adaptively searches for cluster regions in aerial images and resizes them to the appropriate scale for accurate detection, improving efficiency and accuracy. 2. **Refined Regression Loss**: The regression loss is modified using Gaussian Wasserstein distance (GWD) to better handle small objects, and a combination of GWD and $L_1$ loss is proposed to balance performance for both small and large objects. 3. **Improved Detection Head**: The detection head is enhanced with deformable convolutions to refine bounding box regression and a disentangled heatmap branch to improve localization accuracy for different object categories. The effectiveness of YOLC is demonstrated through extensive experiments on two aerial image datasets, VisDrone2019 and UAVDT, showing superior performance compared to state-of-the-art methods in terms of accuracy and efficiency. The paper also includes ablation studies to validate the effectiveness of each component of the proposed framework.