loveda 话题在 GitHub 上没有填写简介，这里按 Star 数汇总最相关的热门仓库。
[IGARSS ORAL 2026] Official implementation of SyntheticGen: “Mitigating Long-Tail Bias via Prompt-Controlled Diffusion Augmentation” — controllable synthetic data generation for long-tailed remote-sensing semantic segmentation with class-ratio and domain control.
HAFR-Net (Int. J. Appl. Earth Obs. Geoinf., under review): identity-initialized refinement of a pretrained Swin-B for VHR aerial image segmentation. LoveDA weights released.
Multi-scale self-attention ConvTransformer (MSSACT-Net) for high-resolution land-cover segmentation: LoveDA benchmark, baseline comparison, module ablation, training-strategy search, and a label-quality ceiling study
Remote sensing image segmentation using SAM, GroundingDINO, and PerSAM with automatic, point, text-prompt, and one-shot workflows, deployed on Hugging Face Spaces using Gradio.
Partial cross-entropy for point-supervised land-cover segmentation: 92% of fully supervised mIoU from 0.3% of the labels, on LoveDA.
PyTorch research project exploring sparse point supervision for LoveDA remote-sensing segmentation with DeepLabV3+ and custom partial cross-entropy.