TinyRXNet: A Local RX-Based Architecture for Accurate Segmentation of Small-Scale Objects

Segmenting tiny objects in high-resolution images, such as pedestrians in aerial views, poses major challenges due to their small size and limited contextual information. Existing segmentation models often fail to capture these fine details accurately. However, such small objects often appear as localized anomalies within the scene—distinct from their surroundings in feature space—which makes local anomaly detection methods like Local Reed-Xialoi (RX) detector, particularly effective for highlighting them. Thus, we introduce TinyRXNet, a novel deep segmentation network that combines SAM-based feature encoding with Local RX blocks, which apply localized anomaly scoring to highlight distinctive regions. This RX-based attention mechanism helps the model focus on areas that deviate from their surroundings, improving segmentation performance. Evaluated on the TinyPedSeg dataset, TinyRXNet outperforms standard baselines, demonstrating the effectiveness of combining deep features with spatial anomaly cues for tiny object segmentation. The codes are available at https://anonymous.4open.science/r/TINYRXNET-693B/