zcablii/LSKNet

What it solves

LSKNet addresses the challenge of detecting objects in remote sensing imagery, where objects can be tiny and require varying amounts of long-range context to be correctly identified. Standard backbones often overlook the unique spatial priors of remote sensing scenarios, leading to potential misdetections of small objects.

How it works

LSKNet is a lightweight foundation backbone that utilizes a Large Selective Kernel mechanism. This allows the network to dynamically adjust its spatial receptive field, enabling it to better model the ranging context required by different types of objects in remote sensing images.

Who it’s for

Researchers and developers working on remote sensing object detection, classification, and semantic segmentation.

Highlights

• State-of-the-art performance: Achieves high mAP on standard benchmarks like DOTA1.0, FAIR1M, and HRSC2016.
• Lightweight design: Provides a foundation backbone that is efficient yet powerful.
• Dynamic receptive field: Uses selective kernels to adapt to the scale and context of objects.
• Integration: Built on MMRotate and compatible with PyTorch 1.6+.