Memory-Constrained Semantic Segmentation for Ultra-High Resolution UAV Imagery

Ultra-high resolution image segmentation poses a formidable challenge for UAVs with limited computation resources. Moreover, with multiple deployed tasks (e.g., mapping, localization, and decision making), the demand for a memory efficient model becomes more urgent. This letter delves into the intricate problem of achieving efficient and effective segmentation of ultra-high resolution UAV imagery, while operating under stringent GPU memory limitation. To address this problem, we propose a GPU memory-efficient and effective framework. Specifically, we introduce a novel and efficient spatial-guided high-resolution query module, which enables our model to effectively infer pixel-wise segmentation results by querying nearest latent embeddings from low-resolution features. Additionally, we present a memory-based interaction scheme with linear complexity to rectify semantic bias beneath the high-resolution spatial guidance via associating cross-image contextual semantics. For evaluation, we perform comprehensive experiments over public benchmarks under both conditions of small and large GPU memory usage limitations. Notably, our model gains around 3% advantage against SOTA in mIoU using comparable memory. Furthermore, we show that our model can be deployed on the embedded platform with less than 8 G memory like Jetson TX2.