RangiLyu/nanodet

Built on a FCOS-style one-stage anchor-free architecture with Generalized Focal Loss, NanoDet-Plus introduces an Assign Guidance Module (AGM) and Dynamic Soft Label Assigner (DSLA) for optimized label assignment in lightweight training, plus Ghost-PAN for efficient multi-scale feature fusion. Deployable across multiple inference frameworks—ncnn, MNN, OpenVINO—with native Android support and PyTorch training via PyTorch Lightning, achieving 34.1 mAP on COCO at competitive latency across CPU, ARM, and mobile platforms.

6,170 stars. No commits in the last 6 months.