Title: Rank-Aware Quantile Activation for Motion-Robust Crop Segmentation in UAV Imagery

Abstract: High-speed Unmanned Aerial Vehicle (UAV) data acquisition often introduces motion blur, which significantly impairs the semantic segmentation of rare, texture-dependent classes that hold high agronomic importance. Traditional Convolutional Neural Networks (CNNs) depend heavily on high-frequency magnitude features; however, these features are obliterated by blur, leading to the statistical disappearance of minority signals. To address this, we introduce Dual Quantile Activation (QAct), a rank-aware module that substitutes magnitude gating with instance-level rank normalization.

We evaluated QAct using the Agriculture-Vision 2021 dataset under both zero-shot and blur-supervised conditions across varying levels of blur severity. The results identify QAct as the primary architectural driver of improvement, demonstrating consistent mean Intersection over Union (mIoU) gains compared to ReLU across all severities and both evaluation regimes. These improvements are particularly pronounced for rare structural and texture-dependent classes. While some dominant classes, such as water and planter skips, exhibited mixed per-class performance during distillation, QAct proved superior in specific contexts. Specifically, zero-shot QAct outperformed distillation-trained ReLU models under moderate blur conditions. Furthermore, the Distill-QAct configuration achieved the highest overall performance across all severity levels, underscoring that rank-aware activation and blur-domain training serve as complementary sources of robustness.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC