<p>Object detection in remote sensing images has become increasingly critical with the rapid development of fields such as unmanned aerial vehicles. However, while rotated bounding box annotations for such images are costly, the use of low-cost point annotations holds great promise. Nevertheless, the inability of point annotations to provide object size and orientation information poses a significant challenge for precise object localization in models. To address this, we propose PSTNet, a framework integrating Template Overlay Learning, Multi-scale Attention Dilated Block (MADB), and Dynamic K-value Sample Assignment. First, category-specific templates are randomly flipped, scaled, and overlaid as pseudo-labels to teach models size and orientation. Second, MADB replaces FPN with dilated convolutions and attention mechanisms to enhance multi-scale feature fusion for small objects. Third, a dynamic K-value strategy leverages classification scores to adaptively assign positive samples, bypassing IoU dependency. Extensive experiments on four datasets show substantial improvements to the baseline.</p>