GaussFocus RPN for Remote Sensing Object Detection

Abstract

Objects in remotesensing imagery are often unevenly distributed and densely clustered, making foreground selection particularly challenging. In two-stage detectors, the Region Proposal Network (RPN) is pivotal, yet proposals from conventional RPNs can be imprecise and redundant, hurting recall and inflating computation.We introduce GaussFocus, a density-guided two-stage detector that builds a unified Gaussian foreground prior from rotated annotations and injects it into proposal generation, feature sampling, and RoI score calibration. Specifically, we derive an image-level Gaussian foreground density map and train a lightweight density branch inside the RPN, where MDeformConv uses regression-guided offsets and density-aware gating to produce features that better align with object geometry and foreground regions. A Gauss Guided Selector further leverages the predicted density map to retain proposals concentrated in foreground-dense areas, reducing background clutter and second-stage computation. In addition, an inference-only Gauss-Guided RoI logit shift calibrates RoI classification scores using the same density prior without retraining, leading to more reliable foreground–background separation and ranking. Across four challenging datasets (including DOTA-v1.0, DOTA-v2.0, and HRSC2016), GaussFocus consistently improves proposal quality and detection accuracy over strong baselines while reducing redundant RoIs and computation.